4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
27 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
28 * Copyright (c) 2014 Integros [integros.com]
31 #include <sys/zfs_context.h>
33 #include <sys/dmu_send.h>
34 #include <sys/dmu_impl.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dmu_tx.h>
42 #include <sys/dmu_zfetch.h>
44 #include <sys/sa_impl.h>
45 #include <sys/zfeature.h>
46 #include <sys/blkptr.h>
47 #include <sys/range_tree.h>
48 #include <sys/callb.h>
51 #include <sys/cityhash.h>
52 #include <sys/spa_impl.h>
56 typedef struct dbuf_stats {
58 * Various statistics about the size of the dbuf cache.
60 kstat_named_t cache_count;
61 kstat_named_t cache_size_bytes;
62 kstat_named_t cache_size_bytes_max;
64 * Statistics regarding the bounds on the dbuf cache size.
66 kstat_named_t cache_target_bytes;
67 kstat_named_t cache_lowater_bytes;
68 kstat_named_t cache_hiwater_bytes;
70 * Total number of dbuf cache evictions that have occurred.
72 kstat_named_t cache_total_evicts;
74 * The distribution of dbuf levels in the dbuf cache and
75 * the total size of all dbufs at each level.
77 kstat_named_t cache_levels[DN_MAX_LEVELS];
78 kstat_named_t cache_levels_bytes[DN_MAX_LEVELS];
80 * Statistics about the dbuf hash table.
82 kstat_named_t hash_hits;
83 kstat_named_t hash_misses;
84 kstat_named_t hash_collisions;
85 kstat_named_t hash_elements;
86 kstat_named_t hash_elements_max;
88 * Number of sublists containing more than one dbuf in the dbuf
89 * hash table. Keep track of the longest hash chain.
91 kstat_named_t hash_chains;
92 kstat_named_t hash_chain_max;
94 * Number of times a dbuf_create() discovers that a dbuf was
95 * already created and in the dbuf hash table.
97 kstat_named_t hash_insert_race;
99 * Statistics about the size of the metadata dbuf cache.
101 kstat_named_t metadata_cache_count;
102 kstat_named_t metadata_cache_size_bytes;
103 kstat_named_t metadata_cache_size_bytes_max;
105 * For diagnostic purposes, this is incremented whenever we can't add
106 * something to the metadata cache because it's full, and instead put
107 * the data in the regular dbuf cache.
109 kstat_named_t metadata_cache_overflow;
112 dbuf_stats_t dbuf_stats = {
113 { "cache_count", KSTAT_DATA_UINT64 },
114 { "cache_size_bytes", KSTAT_DATA_UINT64 },
115 { "cache_size_bytes_max", KSTAT_DATA_UINT64 },
116 { "cache_target_bytes", KSTAT_DATA_UINT64 },
117 { "cache_lowater_bytes", KSTAT_DATA_UINT64 },
118 { "cache_hiwater_bytes", KSTAT_DATA_UINT64 },
119 { "cache_total_evicts", KSTAT_DATA_UINT64 },
120 { { "cache_levels_N", KSTAT_DATA_UINT64 } },
121 { { "cache_levels_bytes_N", KSTAT_DATA_UINT64 } },
122 { "hash_hits", KSTAT_DATA_UINT64 },
123 { "hash_misses", KSTAT_DATA_UINT64 },
124 { "hash_collisions", KSTAT_DATA_UINT64 },
125 { "hash_elements", KSTAT_DATA_UINT64 },
126 { "hash_elements_max", KSTAT_DATA_UINT64 },
127 { "hash_chains", KSTAT_DATA_UINT64 },
128 { "hash_chain_max", KSTAT_DATA_UINT64 },
129 { "hash_insert_race", KSTAT_DATA_UINT64 },
130 { "metadata_cache_count", KSTAT_DATA_UINT64 },
131 { "metadata_cache_size_bytes", KSTAT_DATA_UINT64 },
132 { "metadata_cache_size_bytes_max", KSTAT_DATA_UINT64 },
133 { "metadata_cache_overflow", KSTAT_DATA_UINT64 }
136 #define DBUF_STAT_INCR(stat, val) \
137 atomic_add_64(&dbuf_stats.stat.value.ui64, (val));
138 #define DBUF_STAT_DECR(stat, val) \
139 DBUF_STAT_INCR(stat, -(val));
140 #define DBUF_STAT_BUMP(stat) \
141 DBUF_STAT_INCR(stat, 1);
142 #define DBUF_STAT_BUMPDOWN(stat) \
143 DBUF_STAT_INCR(stat, -1);
144 #define DBUF_STAT_MAX(stat, v) { \
146 while ((v) > (_m = dbuf_stats.stat.value.ui64) && \
147 (_m != atomic_cas_64(&dbuf_stats.stat.value.ui64, _m, (v))))\
151 struct dbuf_hold_impl_data {
152 /* Function arguments */
156 boolean_t dh_fail_sparse;
157 boolean_t dh_fail_uncached;
159 dmu_buf_impl_t **dh_dbp;
160 /* Local variables */
161 dmu_buf_impl_t *dh_db;
162 dmu_buf_impl_t *dh_parent;
165 dbuf_dirty_record_t *dh_dr;
169 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
170 dnode_t *dn, uint8_t level, uint64_t blkid, boolean_t fail_sparse,
171 boolean_t fail_uncached,
172 void *tag, dmu_buf_impl_t **dbp, int depth);
173 static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh);
175 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
176 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
179 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
180 dmu_buf_evict_func_t *evict_func_sync,
181 dmu_buf_evict_func_t *evict_func_async,
182 dmu_buf_t **clear_on_evict_dbufp);
183 #endif /* ! __lint */
186 * Global data structures and functions for the dbuf cache.
188 static kmem_cache_t *dbuf_kmem_cache;
189 static taskq_t *dbu_evict_taskq;
191 static kthread_t *dbuf_cache_evict_thread;
192 static kmutex_t dbuf_evict_lock;
193 static kcondvar_t dbuf_evict_cv;
194 static boolean_t dbuf_evict_thread_exit;
197 * There are two dbuf caches; each dbuf can only be in one of them at a time.
199 * 1. Cache of metadata dbufs, to help make read-heavy administrative commands
200 * from /sbin/zfs run faster. The "metadata cache" specifically stores dbufs
201 * that represent the metadata that describes filesystems/snapshots/
202 * bookmarks/properties/etc. We only evict from this cache when we export a
203 * pool, to short-circuit as much I/O as possible for all administrative
204 * commands that need the metadata. There is no eviction policy for this
205 * cache, because we try to only include types in it which would occupy a
206 * very small amount of space per object but create a large impact on the
207 * performance of these commands. Instead, after it reaches a maximum size
208 * (which should only happen on very small memory systems with a very large
209 * number of filesystem objects), we stop taking new dbufs into the
210 * metadata cache, instead putting them in the normal dbuf cache.
212 * 2. LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
213 * are not currently held but have been recently released. These dbufs
214 * are not eligible for arc eviction until they are aged out of the cache.
215 * Dbufs that are aged out of the cache will be immediately destroyed and
216 * become eligible for arc eviction.
218 * Dbufs are added to these caches once the last hold is released. If a dbuf is
219 * later accessed and still exists in the dbuf cache, then it will be removed
220 * from the cache and later re-added to the head of the cache.
222 * If a given dbuf meets the requirements for the metadata cache, it will go
223 * there, otherwise it will be considered for the generic LRU dbuf cache. The
224 * caches and the refcounts tracking their sizes are stored in an array indexed
225 * by those caches' matching enum values (from dbuf_cached_state_t).
227 typedef struct dbuf_cache {
231 dbuf_cache_t dbuf_caches[DB_CACHE_MAX];
233 /* Size limits for the caches */
234 uint64_t dbuf_cache_max_bytes = 0;
235 uint64_t dbuf_metadata_cache_max_bytes = 0;
236 /* Set the default sizes of the caches to log2 fraction of arc size */
237 int dbuf_cache_shift = 5;
238 int dbuf_metadata_cache_shift = 6;
241 * For diagnostic purposes, this is incremented whenever we can't add
242 * something to the metadata cache because it's full, and instead put
243 * the data in the regular dbuf cache.
245 uint64_t dbuf_metadata_cache_overflow;
248 * The LRU dbuf cache uses a three-stage eviction policy:
249 * - A low water marker designates when the dbuf eviction thread
250 * should stop evicting from the dbuf cache.
251 * - When we reach the maximum size (aka mid water mark), we
252 * signal the eviction thread to run.
253 * - The high water mark indicates when the eviction thread
254 * is unable to keep up with the incoming load and eviction must
255 * happen in the context of the calling thread.
259 * low water mid water hi water
260 * +----------------------------------------+----------+----------+
265 * +----------------------------------------+----------+----------+
267 * evicting eviction directly
270 * The high and low water marks indicate the operating range for the eviction
271 * thread. The low water mark is, by default, 90% of the total size of the
272 * cache and the high water mark is at 110% (both of these percentages can be
273 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
274 * respectively). The eviction thread will try to ensure that the cache remains
275 * within this range by waking up every second and checking if the cache is
276 * above the low water mark. The thread can also be woken up by callers adding
277 * elements into the cache if the cache is larger than the mid water (i.e max
278 * cache size). Once the eviction thread is woken up and eviction is required,
279 * it will continue evicting buffers until it's able to reduce the cache size
280 * to the low water mark. If the cache size continues to grow and hits the high
281 * water mark, then callers adding elments to the cache will begin to evict
282 * directly from the cache until the cache is no longer above the high water
287 * The percentage above and below the maximum cache size.
289 uint_t dbuf_cache_hiwater_pct = 10;
290 uint_t dbuf_cache_lowater_pct = 10;
292 SYSCTL_DECL(_vfs_zfs);
293 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, dbuf_cache_max_bytes, CTLFLAG_RWTUN,
294 &dbuf_cache_max_bytes, 0, "dbuf cache size in bytes");
295 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, dbuf_metadata_cache_max_bytes, CTLFLAG_RWTUN,
296 &dbuf_metadata_cache_max_bytes, 0, "dbuf metadata cache size in bytes");
297 SYSCTL_INT(_vfs_zfs, OID_AUTO, dbuf_cache_shift, CTLFLAG_RDTUN,
298 &dbuf_cache_shift, 0, "dbuf cache size as log2 fraction of ARC");
299 SYSCTL_INT(_vfs_zfs, OID_AUTO, dbuf_metadata_cache_shift, CTLFLAG_RDTUN,
300 &dbuf_metadata_cache_shift, 0,
301 "dbuf metadata cache size as log2 fraction of ARC");
302 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, dbuf_metadata_cache_overflow, CTLFLAG_RD,
303 &dbuf_metadata_cache_overflow, 0, "dbuf metadata cache overflow");
304 SYSCTL_UINT(_vfs_zfs, OID_AUTO, dbuf_cache_hiwater_pct, CTLFLAG_RWTUN,
305 &dbuf_cache_hiwater_pct, 0, "max percents above the dbuf cache size");
306 SYSCTL_UINT(_vfs_zfs, OID_AUTO, dbuf_cache_lowater_pct, CTLFLAG_RWTUN,
307 &dbuf_cache_lowater_pct, 0, "max percents below the dbuf cache size");
311 dbuf_cons(void *vdb, void *unused, int kmflag)
313 dmu_buf_impl_t *db = vdb;
314 bzero(db, sizeof (dmu_buf_impl_t));
316 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
317 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
318 multilist_link_init(&db->db_cache_link);
319 refcount_create(&db->db_holds);
326 dbuf_dest(void *vdb, void *unused)
328 dmu_buf_impl_t *db = vdb;
329 mutex_destroy(&db->db_mtx);
330 cv_destroy(&db->db_changed);
331 ASSERT(!multilist_link_active(&db->db_cache_link));
332 refcount_destroy(&db->db_holds);
336 * dbuf hash table routines
338 static dbuf_hash_table_t dbuf_hash_table;
340 static uint64_t dbuf_hash_count;
343 * We use Cityhash for this. It's fast, and has good hash properties without
344 * requiring any large static buffers.
347 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
349 return (cityhash4((uintptr_t)os, obj, (uint64_t)lvl, blkid));
352 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
353 ((dbuf)->db.db_object == (obj) && \
354 (dbuf)->db_objset == (os) && \
355 (dbuf)->db_level == (level) && \
356 (dbuf)->db_blkid == (blkid))
359 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
361 dbuf_hash_table_t *h = &dbuf_hash_table;
362 uint64_t hv = dbuf_hash(os, obj, level, blkid);
363 uint64_t idx = hv & h->hash_table_mask;
366 mutex_enter(DBUF_HASH_MUTEX(h, idx));
367 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
368 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
369 mutex_enter(&db->db_mtx);
370 if (db->db_state != DB_EVICTING) {
371 mutex_exit(DBUF_HASH_MUTEX(h, idx));
374 mutex_exit(&db->db_mtx);
377 mutex_exit(DBUF_HASH_MUTEX(h, idx));
381 static dmu_buf_impl_t *
382 dbuf_find_bonus(objset_t *os, uint64_t object)
385 dmu_buf_impl_t *db = NULL;
387 if (dnode_hold(os, object, FTAG, &dn) == 0) {
388 rw_enter(&dn->dn_struct_rwlock, RW_READER);
389 if (dn->dn_bonus != NULL) {
391 mutex_enter(&db->db_mtx);
393 rw_exit(&dn->dn_struct_rwlock);
394 dnode_rele(dn, FTAG);
400 * Insert an entry into the hash table. If there is already an element
401 * equal to elem in the hash table, then the already existing element
402 * will be returned and the new element will not be inserted.
403 * Otherwise returns NULL.
405 static dmu_buf_impl_t *
406 dbuf_hash_insert(dmu_buf_impl_t *db)
408 dbuf_hash_table_t *h = &dbuf_hash_table;
409 objset_t *os = db->db_objset;
410 uint64_t obj = db->db.db_object;
411 int level = db->db_level;
412 uint64_t blkid, hv, idx;
416 blkid = db->db_blkid;
417 hv = dbuf_hash(os, obj, level, blkid);
418 idx = hv & h->hash_table_mask;
420 mutex_enter(DBUF_HASH_MUTEX(h, idx));
421 for (dbf = h->hash_table[idx], i = 0; dbf != NULL;
422 dbf = dbf->db_hash_next, i++) {
423 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
424 mutex_enter(&dbf->db_mtx);
425 if (dbf->db_state != DB_EVICTING) {
426 mutex_exit(DBUF_HASH_MUTEX(h, idx));
429 mutex_exit(&dbf->db_mtx);
434 DBUF_STAT_BUMP(hash_collisions);
436 DBUF_STAT_BUMP(hash_chains);
438 DBUF_STAT_MAX(hash_chain_max, i);
441 mutex_enter(&db->db_mtx);
442 db->db_hash_next = h->hash_table[idx];
443 h->hash_table[idx] = db;
444 mutex_exit(DBUF_HASH_MUTEX(h, idx));
445 atomic_inc_64(&dbuf_hash_count);
446 DBUF_STAT_MAX(hash_elements_max, dbuf_hash_count);
452 * Remove an entry from the hash table. It must be in the EVICTING state.
455 dbuf_hash_remove(dmu_buf_impl_t *db)
457 dbuf_hash_table_t *h = &dbuf_hash_table;
459 dmu_buf_impl_t *dbf, **dbp;
461 hv = dbuf_hash(db->db_objset, db->db.db_object,
462 db->db_level, db->db_blkid);
463 idx = hv & h->hash_table_mask;
466 * We mustn't hold db_mtx to maintain lock ordering:
467 * DBUF_HASH_MUTEX > db_mtx.
469 ASSERT(refcount_is_zero(&db->db_holds));
470 ASSERT(db->db_state == DB_EVICTING);
471 ASSERT(!MUTEX_HELD(&db->db_mtx));
473 mutex_enter(DBUF_HASH_MUTEX(h, idx));
474 dbp = &h->hash_table[idx];
475 while ((dbf = *dbp) != db) {
476 dbp = &dbf->db_hash_next;
479 *dbp = db->db_hash_next;
480 db->db_hash_next = NULL;
481 if (h->hash_table[idx] &&
482 h->hash_table[idx]->db_hash_next == NULL)
483 DBUF_STAT_BUMPDOWN(hash_chains);
484 mutex_exit(DBUF_HASH_MUTEX(h, idx));
485 atomic_dec_64(&dbuf_hash_count);
491 } dbvu_verify_type_t;
494 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
499 if (db->db_user == NULL)
502 /* Only data blocks support the attachment of user data. */
503 ASSERT(db->db_level == 0);
505 /* Clients must resolve a dbuf before attaching user data. */
506 ASSERT(db->db.db_data != NULL);
507 ASSERT3U(db->db_state, ==, DB_CACHED);
509 holds = refcount_count(&db->db_holds);
510 if (verify_type == DBVU_EVICTING) {
512 * Immediate eviction occurs when holds == dirtycnt.
513 * For normal eviction buffers, holds is zero on
514 * eviction, except when dbuf_fix_old_data() calls
515 * dbuf_clear_data(). However, the hold count can grow
516 * during eviction even though db_mtx is held (see
517 * dmu_bonus_hold() for an example), so we can only
518 * test the generic invariant that holds >= dirtycnt.
520 ASSERT3U(holds, >=, db->db_dirtycnt);
522 if (db->db_user_immediate_evict == TRUE)
523 ASSERT3U(holds, >=, db->db_dirtycnt);
525 ASSERT3U(holds, >, 0);
531 dbuf_evict_user(dmu_buf_impl_t *db)
533 dmu_buf_user_t *dbu = db->db_user;
535 ASSERT(MUTEX_HELD(&db->db_mtx));
540 dbuf_verify_user(db, DBVU_EVICTING);
544 if (dbu->dbu_clear_on_evict_dbufp != NULL)
545 *dbu->dbu_clear_on_evict_dbufp = NULL;
549 * There are two eviction callbacks - one that we call synchronously
550 * and one that we invoke via a taskq. The async one is useful for
551 * avoiding lock order reversals and limiting stack depth.
553 * Note that if we have a sync callback but no async callback,
554 * it's likely that the sync callback will free the structure
555 * containing the dbu. In that case we need to take care to not
556 * dereference dbu after calling the sync evict func.
558 boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
560 if (dbu->dbu_evict_func_sync != NULL)
561 dbu->dbu_evict_func_sync(dbu);
564 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
565 dbu, 0, &dbu->dbu_tqent);
570 dbuf_is_metadata(dmu_buf_impl_t *db)
572 if (db->db_level > 0) {
575 boolean_t is_metadata;
578 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
581 return (is_metadata);
586 * This returns whether this dbuf should be stored in the metadata cache, which
587 * is based on whether it's from one of the dnode types that store data related
588 * to traversing dataset hierarchies.
591 dbuf_include_in_metadata_cache(dmu_buf_impl_t *db)
594 dmu_object_type_t type = DB_DNODE(db)->dn_type;
597 /* Check if this dbuf is one of the types we care about */
598 if (DMU_OT_IS_METADATA_CACHED(type)) {
599 /* If we hit this, then we set something up wrong in dmu_ot */
600 ASSERT(DMU_OT_IS_METADATA(type));
603 * Sanity check for small-memory systems: don't allocate too
604 * much memory for this purpose.
606 if (refcount_count(&dbuf_caches[DB_DBUF_METADATA_CACHE].size) >
607 dbuf_metadata_cache_max_bytes) {
608 dbuf_metadata_cache_overflow++;
609 DTRACE_PROBE1(dbuf__metadata__cache__overflow,
610 dmu_buf_impl_t *, db);
621 * This function *must* return indices evenly distributed between all
622 * sublists of the multilist. This is needed due to how the dbuf eviction
623 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
624 * distributed between all sublists and uses this assumption when
625 * deciding which sublist to evict from and how much to evict from it.
628 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
630 dmu_buf_impl_t *db = obj;
633 * The assumption here, is the hash value for a given
634 * dmu_buf_impl_t will remain constant throughout it's lifetime
635 * (i.e. it's objset, object, level and blkid fields don't change).
636 * Thus, we don't need to store the dbuf's sublist index
637 * on insertion, as this index can be recalculated on removal.
639 * Also, the low order bits of the hash value are thought to be
640 * distributed evenly. Otherwise, in the case that the multilist
641 * has a power of two number of sublists, each sublists' usage
642 * would not be evenly distributed.
644 return (dbuf_hash(db->db_objset, db->db.db_object,
645 db->db_level, db->db_blkid) %
646 multilist_get_num_sublists(ml));
649 static inline unsigned long
650 dbuf_cache_target_bytes(void)
652 return MIN(dbuf_cache_max_bytes,
653 arc_max_bytes() >> dbuf_cache_shift);
656 static inline uint64_t
657 dbuf_cache_hiwater_bytes(void)
659 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
660 return (dbuf_cache_target +
661 (dbuf_cache_target * dbuf_cache_hiwater_pct) / 100);
664 static inline uint64_t
665 dbuf_cache_lowater_bytes(void)
667 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
668 return (dbuf_cache_target -
669 (dbuf_cache_target * dbuf_cache_lowater_pct) / 100);
672 static inline boolean_t
673 dbuf_cache_above_hiwater(void)
675 return (refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) >
676 dbuf_cache_hiwater_bytes());
679 static inline boolean_t
680 dbuf_cache_above_lowater(void)
682 return (refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) >
683 dbuf_cache_lowater_bytes());
687 * Evict the oldest eligible dbuf from the dbuf cache.
692 int idx = multilist_get_random_index(dbuf_caches[DB_DBUF_CACHE].cache);
693 multilist_sublist_t *mls = multilist_sublist_lock(
694 dbuf_caches[DB_DBUF_CACHE].cache, idx);
696 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
698 dmu_buf_impl_t *db = multilist_sublist_tail(mls);
699 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
700 db = multilist_sublist_prev(mls, db);
703 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
704 multilist_sublist_t *, mls);
707 multilist_sublist_remove(mls, db);
708 multilist_sublist_unlock(mls);
709 (void) refcount_remove_many(&dbuf_caches[DB_DBUF_CACHE].size,
711 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
712 DBUF_STAT_BUMPDOWN(cache_count);
713 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
715 ASSERT3U(db->db_caching_status, ==, DB_DBUF_CACHE);
716 db->db_caching_status = DB_NO_CACHE;
718 DBUF_STAT_MAX(cache_size_bytes_max,
719 refcount_count(&dbuf_caches[DB_DBUF_CACHE].size));
720 DBUF_STAT_BUMP(cache_total_evicts);
722 multilist_sublist_unlock(mls);
727 * The dbuf evict thread is responsible for aging out dbufs from the
728 * cache. Once the cache has reached it's maximum size, dbufs are removed
729 * and destroyed. The eviction thread will continue running until the size
730 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
731 * out of the cache it is destroyed and becomes eligible for arc eviction.
735 dbuf_evict_thread(void *unused __unused)
739 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
741 mutex_enter(&dbuf_evict_lock);
742 while (!dbuf_evict_thread_exit) {
743 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
744 CALLB_CPR_SAFE_BEGIN(&cpr);
745 (void) cv_timedwait_hires(&dbuf_evict_cv,
746 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
747 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
749 mutex_exit(&dbuf_evict_lock);
752 * Keep evicting as long as we're above the low water mark
753 * for the cache. We do this without holding the locks to
754 * minimize lock contention.
756 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
760 mutex_enter(&dbuf_evict_lock);
763 dbuf_evict_thread_exit = B_FALSE;
764 cv_broadcast(&dbuf_evict_cv);
765 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
770 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
771 * If the dbuf cache is at its high water mark, then evict a dbuf from the
772 * dbuf cache using the callers context.
775 dbuf_evict_notify(void)
778 * We check if we should evict without holding the dbuf_evict_lock,
779 * because it's OK to occasionally make the wrong decision here,
780 * and grabbing the lock results in massive lock contention.
782 if (refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) >
783 dbuf_cache_max_bytes) {
784 if (dbuf_cache_above_hiwater())
786 cv_signal(&dbuf_evict_cv);
791 dbuf_kstat_update(kstat_t *ksp, int rw)
793 dbuf_stats_t *ds = ksp->ks_data;
795 if (rw == KSTAT_WRITE) {
796 return (SET_ERROR(EACCES));
798 ds->metadata_cache_size_bytes.value.ui64 =
799 refcount_count(&dbuf_caches[DB_DBUF_METADATA_CACHE].size);
800 ds->cache_size_bytes.value.ui64 =
801 refcount_count(&dbuf_caches[DB_DBUF_CACHE].size);
802 ds->cache_target_bytes.value.ui64 = dbuf_cache_target_bytes();
803 ds->cache_hiwater_bytes.value.ui64 = dbuf_cache_hiwater_bytes();
804 ds->cache_lowater_bytes.value.ui64 = dbuf_cache_lowater_bytes();
805 ds->hash_elements.value.ui64 = dbuf_hash_count;
814 uint64_t hsize = 1ULL << 16;
815 dbuf_hash_table_t *h = &dbuf_hash_table;
819 * The hash table is big enough to fill all of physical memory
820 * with an average 4K block size. The table will take up
821 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
823 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE)
827 h->hash_table_mask = hsize - 1;
828 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
829 if (h->hash_table == NULL) {
830 /* XXX - we should really return an error instead of assert */
831 ASSERT(hsize > (1ULL << 10));
836 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
837 sizeof (dmu_buf_impl_t),
838 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
840 for (i = 0; i < DBUF_MUTEXES; i++)
841 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
845 * Setup the parameters for the dbuf caches. We set the sizes of the
846 * dbuf cache and the metadata cache to 1/32nd and 1/16th (default)
847 * of the size of the ARC, respectively. If the values are set in
848 * /etc/system and they're not greater than the size of the ARC, then
849 * we honor that value.
851 if (dbuf_cache_max_bytes == 0 ||
852 dbuf_cache_max_bytes >= arc_max_bytes()) {
853 dbuf_cache_max_bytes = arc_max_bytes() >> dbuf_cache_shift;
855 if (dbuf_metadata_cache_max_bytes == 0 ||
856 dbuf_metadata_cache_max_bytes >= arc_max_bytes()) {
857 dbuf_metadata_cache_max_bytes =
858 arc_max_bytes() >> dbuf_metadata_cache_shift;
862 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
863 * configuration is not required.
865 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
867 for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) {
868 dbuf_caches[dcs].cache =
869 multilist_create(sizeof (dmu_buf_impl_t),
870 offsetof(dmu_buf_impl_t, db_cache_link),
871 dbuf_cache_multilist_index_func);
872 refcount_create(&dbuf_caches[dcs].size);
875 dbuf_evict_thread_exit = B_FALSE;
876 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
877 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
878 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
879 NULL, 0, &p0, TS_RUN, minclsyspri);
883 * XXX FreeBSD's SPL lacks KSTAT_TYPE_NAMED support - TODO
885 dbuf_ksp = kstat_create("zfs", 0, "dbufstats", "misc",
886 KSTAT_TYPE_NAMED, sizeof (dbuf_stats) / sizeof (kstat_named_t),
888 if (dbuf_ksp != NULL) {
889 dbuf_ksp->ks_data = &dbuf_stats;
890 dbuf_ksp->ks_update = dbuf_kstat_update;
891 kstat_install(dbuf_ksp);
893 for (i = 0; i < DN_MAX_LEVELS; i++) {
894 snprintf(dbuf_stats.cache_levels[i].name,
895 KSTAT_STRLEN, "cache_level_%d", i);
896 dbuf_stats.cache_levels[i].data_type =
898 snprintf(dbuf_stats.cache_levels_bytes[i].name,
899 KSTAT_STRLEN, "cache_level_%d_bytes", i);
900 dbuf_stats.cache_levels_bytes[i].data_type =
910 dbuf_hash_table_t *h = &dbuf_hash_table;
913 dbuf_stats_destroy();
915 for (i = 0; i < DBUF_MUTEXES; i++)
916 mutex_destroy(&h->hash_mutexes[i]);
917 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
918 kmem_cache_destroy(dbuf_kmem_cache);
919 taskq_destroy(dbu_evict_taskq);
921 mutex_enter(&dbuf_evict_lock);
922 dbuf_evict_thread_exit = B_TRUE;
923 while (dbuf_evict_thread_exit) {
924 cv_signal(&dbuf_evict_cv);
925 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
927 mutex_exit(&dbuf_evict_lock);
929 mutex_destroy(&dbuf_evict_lock);
930 cv_destroy(&dbuf_evict_cv);
932 for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) {
933 refcount_destroy(&dbuf_caches[dcs].size);
934 multilist_destroy(dbuf_caches[dcs].cache);
937 if (dbuf_ksp != NULL) {
938 kstat_delete(dbuf_ksp);
949 dbuf_verify(dmu_buf_impl_t *db)
952 dbuf_dirty_record_t *dr;
954 ASSERT(MUTEX_HELD(&db->db_mtx));
956 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
959 ASSERT(db->db_objset != NULL);
963 ASSERT(db->db_parent == NULL);
964 ASSERT(db->db_blkptr == NULL);
966 ASSERT3U(db->db.db_object, ==, dn->dn_object);
967 ASSERT3P(db->db_objset, ==, dn->dn_objset);
968 ASSERT3U(db->db_level, <, dn->dn_nlevels);
969 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
970 db->db_blkid == DMU_SPILL_BLKID ||
971 !avl_is_empty(&dn->dn_dbufs));
973 if (db->db_blkid == DMU_BONUS_BLKID) {
975 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
976 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
977 } else if (db->db_blkid == DMU_SPILL_BLKID) {
979 ASSERT0(db->db.db_offset);
981 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
984 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
985 ASSERT(dr->dr_dbuf == db);
987 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
988 ASSERT(dr->dr_dbuf == db);
991 * We can't assert that db_size matches dn_datablksz because it
992 * can be momentarily different when another thread is doing
995 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
996 dr = db->db_data_pending;
998 * It should only be modified in syncing context, so
999 * make sure we only have one copy of the data.
1001 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
1004 /* verify db->db_blkptr */
1005 if (db->db_blkptr) {
1006 if (db->db_parent == dn->dn_dbuf) {
1007 /* db is pointed to by the dnode */
1008 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
1009 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
1010 ASSERT(db->db_parent == NULL);
1012 ASSERT(db->db_parent != NULL);
1013 if (db->db_blkid != DMU_SPILL_BLKID)
1014 ASSERT3P(db->db_blkptr, ==,
1015 &dn->dn_phys->dn_blkptr[db->db_blkid]);
1017 /* db is pointed to by an indirect block */
1018 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
1019 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
1020 ASSERT3U(db->db_parent->db.db_object, ==,
1023 * dnode_grow_indblksz() can make this fail if we don't
1024 * have the struct_rwlock. XXX indblksz no longer
1025 * grows. safe to do this now?
1027 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1028 ASSERT3P(db->db_blkptr, ==,
1029 ((blkptr_t *)db->db_parent->db.db_data +
1030 db->db_blkid % epb));
1034 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
1035 (db->db_buf == NULL || db->db_buf->b_data) &&
1036 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
1037 db->db_state != DB_FILL && !dn->dn_free_txg) {
1039 * If the blkptr isn't set but they have nonzero data,
1040 * it had better be dirty, otherwise we'll lose that
1041 * data when we evict this buffer.
1043 * There is an exception to this rule for indirect blocks; in
1044 * this case, if the indirect block is a hole, we fill in a few
1045 * fields on each of the child blocks (importantly, birth time)
1046 * to prevent hole birth times from being lost when you
1047 * partially fill in a hole.
1049 if (db->db_dirtycnt == 0) {
1050 if (db->db_level == 0) {
1051 uint64_t *buf = db->db.db_data;
1054 for (i = 0; i < db->db.db_size >> 3; i++) {
1055 ASSERT(buf[i] == 0);
1058 blkptr_t *bps = db->db.db_data;
1059 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
1062 * We want to verify that all the blkptrs in the
1063 * indirect block are holes, but we may have
1064 * automatically set up a few fields for them.
1065 * We iterate through each blkptr and verify
1066 * they only have those fields set.
1069 i < db->db.db_size / sizeof (blkptr_t);
1071 blkptr_t *bp = &bps[i];
1072 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1075 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
1076 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
1077 DVA_IS_EMPTY(&bp->blk_dva[2]));
1078 ASSERT0(bp->blk_fill);
1079 ASSERT0(bp->blk_pad[0]);
1080 ASSERT0(bp->blk_pad[1]);
1081 ASSERT(!BP_IS_EMBEDDED(bp));
1082 ASSERT(BP_IS_HOLE(bp));
1083 ASSERT0(bp->blk_phys_birth);
1093 dbuf_clear_data(dmu_buf_impl_t *db)
1095 ASSERT(MUTEX_HELD(&db->db_mtx));
1096 dbuf_evict_user(db);
1097 ASSERT3P(db->db_buf, ==, NULL);
1098 db->db.db_data = NULL;
1099 if (db->db_state != DB_NOFILL)
1100 db->db_state = DB_UNCACHED;
1104 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
1106 ASSERT(MUTEX_HELD(&db->db_mtx));
1107 ASSERT(buf != NULL);
1110 ASSERT(buf->b_data != NULL);
1111 db->db.db_data = buf->b_data;
1115 * Loan out an arc_buf for read. Return the loaned arc_buf.
1118 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
1122 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1123 mutex_enter(&db->db_mtx);
1124 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
1125 int blksz = db->db.db_size;
1126 spa_t *spa = db->db_objset->os_spa;
1128 mutex_exit(&db->db_mtx);
1129 abuf = arc_loan_buf(spa, B_FALSE, blksz);
1130 bcopy(db->db.db_data, abuf->b_data, blksz);
1133 arc_loan_inuse_buf(abuf, db);
1135 dbuf_clear_data(db);
1136 mutex_exit(&db->db_mtx);
1142 * Calculate which level n block references the data at the level 0 offset
1146 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
1148 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
1150 * The level n blkid is equal to the level 0 blkid divided by
1151 * the number of level 0s in a level n block.
1153 * The level 0 blkid is offset >> datablkshift =
1154 * offset / 2^datablkshift.
1156 * The number of level 0s in a level n is the number of block
1157 * pointers in an indirect block, raised to the power of level.
1158 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1159 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1161 * Thus, the level n blkid is: offset /
1162 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
1163 * = offset / 2^(datablkshift + level *
1164 * (indblkshift - SPA_BLKPTRSHIFT))
1165 * = offset >> (datablkshift + level *
1166 * (indblkshift - SPA_BLKPTRSHIFT))
1168 return (offset >> (dn->dn_datablkshift + level *
1169 (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
1171 ASSERT3U(offset, <, dn->dn_datablksz);
1177 dbuf_read_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1178 arc_buf_t *buf, void *vdb)
1180 dmu_buf_impl_t *db = vdb;
1182 mutex_enter(&db->db_mtx);
1183 ASSERT3U(db->db_state, ==, DB_READ);
1185 * All reads are synchronous, so we must have a hold on the dbuf
1187 ASSERT(refcount_count(&db->db_holds) > 0);
1188 ASSERT(db->db_buf == NULL);
1189 ASSERT(db->db.db_data == NULL);
1192 ASSERT(zio == NULL || zio->io_error != 0);
1193 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1194 ASSERT3P(db->db_buf, ==, NULL);
1195 db->db_state = DB_UNCACHED;
1196 } else if (db->db_level == 0 && db->db_freed_in_flight) {
1197 /* freed in flight */
1198 ASSERT(zio == NULL || zio->io_error == 0);
1200 buf = arc_alloc_buf(db->db_objset->os_spa,
1201 db, DBUF_GET_BUFC_TYPE(db), db->db.db_size);
1203 arc_release(buf, db);
1204 bzero(buf->b_data, db->db.db_size);
1205 arc_buf_freeze(buf);
1206 db->db_freed_in_flight = FALSE;
1207 dbuf_set_data(db, buf);
1208 db->db_state = DB_CACHED;
1211 ASSERT(zio == NULL || zio->io_error == 0);
1212 dbuf_set_data(db, buf);
1213 db->db_state = DB_CACHED;
1215 cv_broadcast(&db->db_changed);
1216 dbuf_rele_and_unlock(db, NULL, B_FALSE);
1220 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1223 zbookmark_phys_t zb;
1224 arc_flags_t aflags = ARC_FLAG_NOWAIT;
1228 ASSERT(!refcount_is_zero(&db->db_holds));
1229 /* We need the struct_rwlock to prevent db_blkptr from changing. */
1230 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1231 ASSERT(MUTEX_HELD(&db->db_mtx));
1232 ASSERT(db->db_state == DB_UNCACHED);
1233 ASSERT(db->db_buf == NULL);
1235 if (db->db_blkid == DMU_BONUS_BLKID) {
1237 * The bonus length stored in the dnode may be less than
1238 * the maximum available space in the bonus buffer.
1240 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
1241 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1243 ASSERT3U(bonuslen, <=, db->db.db_size);
1244 db->db.db_data = zio_buf_alloc(max_bonuslen);
1245 arc_space_consume(max_bonuslen, ARC_SPACE_BONUS);
1246 if (bonuslen < max_bonuslen)
1247 bzero(db->db.db_data, max_bonuslen);
1249 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
1251 db->db_state = DB_CACHED;
1252 mutex_exit(&db->db_mtx);
1257 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1258 * processes the delete record and clears the bp while we are waiting
1259 * for the dn_mtx (resulting in a "no" from block_freed).
1261 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
1262 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
1263 BP_IS_HOLE(db->db_blkptr)))) {
1264 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1266 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
1268 bzero(db->db.db_data, db->db.db_size);
1270 if (db->db_blkptr != NULL && db->db_level > 0 &&
1271 BP_IS_HOLE(db->db_blkptr) &&
1272 db->db_blkptr->blk_birth != 0) {
1273 blkptr_t *bps = db->db.db_data;
1274 for (int i = 0; i < ((1 <<
1275 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
1277 blkptr_t *bp = &bps[i];
1278 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
1279 1 << dn->dn_indblkshift);
1281 BP_GET_LEVEL(db->db_blkptr) == 1 ?
1283 BP_GET_LSIZE(db->db_blkptr));
1284 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
1286 BP_GET_LEVEL(db->db_blkptr) - 1);
1287 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
1291 db->db_state = DB_CACHED;
1292 mutex_exit(&db->db_mtx);
1298 db->db_state = DB_READ;
1299 mutex_exit(&db->db_mtx);
1301 if (DBUF_IS_L2CACHEABLE(db))
1302 aflags |= ARC_FLAG_L2CACHE;
1304 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1305 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1306 db->db.db_object, db->db_level, db->db_blkid);
1308 dbuf_add_ref(db, NULL);
1310 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1311 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1312 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
1317 * This is our just-in-time copy function. It makes a copy of buffers that
1318 * have been modified in a previous transaction group before we access them in
1319 * the current active group.
1321 * This function is used in three places: when we are dirtying a buffer for the
1322 * first time in a txg, when we are freeing a range in a dnode that includes
1323 * this buffer, and when we are accessing a buffer which was received compressed
1324 * and later referenced in a WRITE_BYREF record.
1326 * Note that when we are called from dbuf_free_range() we do not put a hold on
1327 * the buffer, we just traverse the active dbuf list for the dnode.
1330 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1332 dbuf_dirty_record_t *dr = db->db_last_dirty;
1334 ASSERT(MUTEX_HELD(&db->db_mtx));
1335 ASSERT(db->db.db_data != NULL);
1336 ASSERT(db->db_level == 0);
1337 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1340 (dr->dt.dl.dr_data !=
1341 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1345 * If the last dirty record for this dbuf has not yet synced
1346 * and its referencing the dbuf data, either:
1347 * reset the reference to point to a new copy,
1348 * or (if there a no active holders)
1349 * just null out the current db_data pointer.
1351 ASSERT(dr->dr_txg >= txg - 2);
1352 if (db->db_blkid == DMU_BONUS_BLKID) {
1353 /* Note that the data bufs here are zio_bufs */
1354 dnode_t *dn = DB_DNODE(db);
1355 int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1356 dr->dt.dl.dr_data = zio_buf_alloc(bonuslen);
1357 arc_space_consume(bonuslen, ARC_SPACE_BONUS);
1358 bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen);
1359 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1360 int size = arc_buf_size(db->db_buf);
1361 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1362 spa_t *spa = db->db_objset->os_spa;
1363 enum zio_compress compress_type =
1364 arc_get_compression(db->db_buf);
1366 if (compress_type == ZIO_COMPRESS_OFF) {
1367 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1369 ASSERT3U(type, ==, ARC_BUFC_DATA);
1370 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1371 size, arc_buf_lsize(db->db_buf), compress_type);
1373 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1376 dbuf_clear_data(db);
1381 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1388 * We don't have to hold the mutex to check db_state because it
1389 * can't be freed while we have a hold on the buffer.
1391 ASSERT(!refcount_is_zero(&db->db_holds));
1393 if (db->db_state == DB_NOFILL)
1394 return (SET_ERROR(EIO));
1398 if ((flags & DB_RF_HAVESTRUCT) == 0)
1399 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1401 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1402 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1403 DBUF_IS_CACHEABLE(db);
1405 mutex_enter(&db->db_mtx);
1406 if (db->db_state == DB_CACHED) {
1408 * If the arc buf is compressed, we need to decompress it to
1409 * read the data. This could happen during the "zfs receive" of
1410 * a stream which is compressed and deduplicated.
1412 if (db->db_buf != NULL &&
1413 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) {
1414 dbuf_fix_old_data(db,
1415 spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1416 err = arc_decompress(db->db_buf);
1417 dbuf_set_data(db, db->db_buf);
1419 mutex_exit(&db->db_mtx);
1421 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1422 if ((flags & DB_RF_HAVESTRUCT) == 0)
1423 rw_exit(&dn->dn_struct_rwlock);
1425 DBUF_STAT_BUMP(hash_hits);
1426 } else if (db->db_state == DB_UNCACHED) {
1427 spa_t *spa = dn->dn_objset->os_spa;
1428 boolean_t need_wait = B_FALSE;
1431 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1432 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1435 dbuf_read_impl(db, zio, flags);
1437 /* dbuf_read_impl has dropped db_mtx for us */
1440 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1442 if ((flags & DB_RF_HAVESTRUCT) == 0)
1443 rw_exit(&dn->dn_struct_rwlock);
1445 DBUF_STAT_BUMP(hash_misses);
1448 err = zio_wait(zio);
1451 * Another reader came in while the dbuf was in flight
1452 * between UNCACHED and CACHED. Either a writer will finish
1453 * writing the buffer (sending the dbuf to CACHED) or the
1454 * first reader's request will reach the read_done callback
1455 * and send the dbuf to CACHED. Otherwise, a failure
1456 * occurred and the dbuf went to UNCACHED.
1458 mutex_exit(&db->db_mtx);
1460 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1461 if ((flags & DB_RF_HAVESTRUCT) == 0)
1462 rw_exit(&dn->dn_struct_rwlock);
1464 DBUF_STAT_BUMP(hash_misses);
1466 /* Skip the wait per the caller's request. */
1467 mutex_enter(&db->db_mtx);
1468 if ((flags & DB_RF_NEVERWAIT) == 0) {
1469 while (db->db_state == DB_READ ||
1470 db->db_state == DB_FILL) {
1471 ASSERT(db->db_state == DB_READ ||
1472 (flags & DB_RF_HAVESTRUCT) == 0);
1473 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1475 cv_wait(&db->db_changed, &db->db_mtx);
1477 if (db->db_state == DB_UNCACHED)
1478 err = SET_ERROR(EIO);
1480 mutex_exit(&db->db_mtx);
1487 dbuf_noread(dmu_buf_impl_t *db)
1489 ASSERT(!refcount_is_zero(&db->db_holds));
1490 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1491 mutex_enter(&db->db_mtx);
1492 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1493 cv_wait(&db->db_changed, &db->db_mtx);
1494 if (db->db_state == DB_UNCACHED) {
1495 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1496 spa_t *spa = db->db_objset->os_spa;
1498 ASSERT(db->db_buf == NULL);
1499 ASSERT(db->db.db_data == NULL);
1500 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1501 db->db_state = DB_FILL;
1502 } else if (db->db_state == DB_NOFILL) {
1503 dbuf_clear_data(db);
1505 ASSERT3U(db->db_state, ==, DB_CACHED);
1507 mutex_exit(&db->db_mtx);
1511 dbuf_unoverride(dbuf_dirty_record_t *dr)
1513 dmu_buf_impl_t *db = dr->dr_dbuf;
1514 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1515 uint64_t txg = dr->dr_txg;
1517 ASSERT(MUTEX_HELD(&db->db_mtx));
1519 * This assert is valid because dmu_sync() expects to be called by
1520 * a zilog's get_data while holding a range lock. This call only
1521 * comes from dbuf_dirty() callers who must also hold a range lock.
1523 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1524 ASSERT(db->db_level == 0);
1526 if (db->db_blkid == DMU_BONUS_BLKID ||
1527 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1530 ASSERT(db->db_data_pending != dr);
1532 /* free this block */
1533 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1534 zio_free(db->db_objset->os_spa, txg, bp);
1536 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1537 dr->dt.dl.dr_nopwrite = B_FALSE;
1540 * Release the already-written buffer, so we leave it in
1541 * a consistent dirty state. Note that all callers are
1542 * modifying the buffer, so they will immediately do
1543 * another (redundant) arc_release(). Therefore, leave
1544 * the buf thawed to save the effort of freezing &
1545 * immediately re-thawing it.
1547 arc_release(dr->dt.dl.dr_data, db);
1551 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1552 * data blocks in the free range, so that any future readers will find
1556 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1559 dmu_buf_impl_t db_search;
1560 dmu_buf_impl_t *db, *db_next;
1561 uint64_t txg = tx->tx_txg;
1564 if (end_blkid > dn->dn_maxblkid &&
1565 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1566 end_blkid = dn->dn_maxblkid;
1567 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1569 db_search.db_level = 0;
1570 db_search.db_blkid = start_blkid;
1571 db_search.db_state = DB_SEARCH;
1573 mutex_enter(&dn->dn_dbufs_mtx);
1574 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1575 ASSERT3P(db, ==, NULL);
1577 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1579 for (; db != NULL; db = db_next) {
1580 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1581 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1583 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1586 ASSERT3U(db->db_blkid, >=, start_blkid);
1588 /* found a level 0 buffer in the range */
1589 mutex_enter(&db->db_mtx);
1590 if (dbuf_undirty(db, tx)) {
1591 /* mutex has been dropped and dbuf destroyed */
1595 if (db->db_state == DB_UNCACHED ||
1596 db->db_state == DB_NOFILL ||
1597 db->db_state == DB_EVICTING) {
1598 ASSERT(db->db.db_data == NULL);
1599 mutex_exit(&db->db_mtx);
1602 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1603 /* will be handled in dbuf_read_done or dbuf_rele */
1604 db->db_freed_in_flight = TRUE;
1605 mutex_exit(&db->db_mtx);
1608 if (refcount_count(&db->db_holds) == 0) {
1613 /* The dbuf is referenced */
1615 if (db->db_last_dirty != NULL) {
1616 dbuf_dirty_record_t *dr = db->db_last_dirty;
1618 if (dr->dr_txg == txg) {
1620 * This buffer is "in-use", re-adjust the file
1621 * size to reflect that this buffer may
1622 * contain new data when we sync.
1624 if (db->db_blkid != DMU_SPILL_BLKID &&
1625 db->db_blkid > dn->dn_maxblkid)
1626 dn->dn_maxblkid = db->db_blkid;
1627 dbuf_unoverride(dr);
1630 * This dbuf is not dirty in the open context.
1631 * Either uncache it (if its not referenced in
1632 * the open context) or reset its contents to
1635 dbuf_fix_old_data(db, txg);
1638 /* clear the contents if its cached */
1639 if (db->db_state == DB_CACHED) {
1640 ASSERT(db->db.db_data != NULL);
1641 arc_release(db->db_buf, db);
1642 bzero(db->db.db_data, db->db.db_size);
1643 arc_buf_freeze(db->db_buf);
1646 mutex_exit(&db->db_mtx);
1648 mutex_exit(&dn->dn_dbufs_mtx);
1652 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1654 arc_buf_t *buf, *obuf;
1655 int osize = db->db.db_size;
1656 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1659 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1664 /* XXX does *this* func really need the lock? */
1665 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1668 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1669 * is OK, because there can be no other references to the db
1670 * when we are changing its size, so no concurrent DB_FILL can
1674 * XXX we should be doing a dbuf_read, checking the return
1675 * value and returning that up to our callers
1677 dmu_buf_will_dirty(&db->db, tx);
1679 /* create the data buffer for the new block */
1680 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1682 /* copy old block data to the new block */
1684 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1685 /* zero the remainder */
1687 bzero((uint8_t *)buf->b_data + osize, size - osize);
1689 mutex_enter(&db->db_mtx);
1690 dbuf_set_data(db, buf);
1691 arc_buf_destroy(obuf, db);
1692 db->db.db_size = size;
1694 if (db->db_level == 0) {
1695 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1696 db->db_last_dirty->dt.dl.dr_data = buf;
1698 mutex_exit(&db->db_mtx);
1700 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1705 dbuf_release_bp(dmu_buf_impl_t *db)
1707 objset_t *os = db->db_objset;
1709 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1710 ASSERT(arc_released(os->os_phys_buf) ||
1711 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1712 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1714 (void) arc_release(db->db_buf, db);
1718 * We already have a dirty record for this TXG, and we are being
1722 dbuf_redirty(dbuf_dirty_record_t *dr)
1724 dmu_buf_impl_t *db = dr->dr_dbuf;
1726 ASSERT(MUTEX_HELD(&db->db_mtx));
1728 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1730 * If this buffer has already been written out,
1731 * we now need to reset its state.
1733 dbuf_unoverride(dr);
1734 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1735 db->db_state != DB_NOFILL) {
1736 /* Already released on initial dirty, so just thaw. */
1737 ASSERT(arc_released(db->db_buf));
1738 arc_buf_thaw(db->db_buf);
1743 dbuf_dirty_record_t *
1744 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1748 dbuf_dirty_record_t **drp, *dr;
1749 int drop_struct_lock = FALSE;
1750 int txgoff = tx->tx_txg & TXG_MASK;
1752 ASSERT(tx->tx_txg != 0);
1753 ASSERT(!refcount_is_zero(&db->db_holds));
1754 DMU_TX_DIRTY_BUF(tx, db);
1759 * Shouldn't dirty a regular buffer in syncing context. Private
1760 * objects may be dirtied in syncing context, but only if they
1761 * were already pre-dirtied in open context.
1764 if (dn->dn_objset->os_dsl_dataset != NULL) {
1765 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1768 ASSERT(!dmu_tx_is_syncing(tx) ||
1769 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1770 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1771 dn->dn_objset->os_dsl_dataset == NULL);
1772 if (dn->dn_objset->os_dsl_dataset != NULL)
1773 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1776 * We make this assert for private objects as well, but after we
1777 * check if we're already dirty. They are allowed to re-dirty
1778 * in syncing context.
1780 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1781 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1782 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1784 mutex_enter(&db->db_mtx);
1786 * XXX make this true for indirects too? The problem is that
1787 * transactions created with dmu_tx_create_assigned() from
1788 * syncing context don't bother holding ahead.
1790 ASSERT(db->db_level != 0 ||
1791 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1792 db->db_state == DB_NOFILL);
1794 mutex_enter(&dn->dn_mtx);
1796 * Don't set dirtyctx to SYNC if we're just modifying this as we
1797 * initialize the objset.
1799 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1800 if (dn->dn_objset->os_dsl_dataset != NULL) {
1801 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1804 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1805 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1806 DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1807 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1808 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1810 if (dn->dn_objset->os_dsl_dataset != NULL) {
1811 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1815 mutex_exit(&dn->dn_mtx);
1817 if (db->db_blkid == DMU_SPILL_BLKID)
1818 dn->dn_have_spill = B_TRUE;
1821 * If this buffer is already dirty, we're done.
1823 drp = &db->db_last_dirty;
1824 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1825 db->db.db_object == DMU_META_DNODE_OBJECT);
1826 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1828 if (dr && dr->dr_txg == tx->tx_txg) {
1832 mutex_exit(&db->db_mtx);
1837 * Only valid if not already dirty.
1839 ASSERT(dn->dn_object == 0 ||
1840 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1841 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1843 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1846 * We should only be dirtying in syncing context if it's the
1847 * mos or we're initializing the os or it's a special object.
1848 * However, we are allowed to dirty in syncing context provided
1849 * we already dirtied it in open context. Hence we must make
1850 * this assertion only if we're not already dirty.
1853 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1855 if (dn->dn_objset->os_dsl_dataset != NULL)
1856 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1857 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1858 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1859 if (dn->dn_objset->os_dsl_dataset != NULL)
1860 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1862 ASSERT(db->db.db_size != 0);
1864 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1866 if (db->db_blkid != DMU_BONUS_BLKID) {
1867 dmu_objset_willuse_space(os, db->db.db_size, tx);
1871 * If this buffer is dirty in an old transaction group we need
1872 * to make a copy of it so that the changes we make in this
1873 * transaction group won't leak out when we sync the older txg.
1875 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1876 list_link_init(&dr->dr_dirty_node);
1877 if (db->db_level == 0) {
1878 void *data_old = db->db_buf;
1880 if (db->db_state != DB_NOFILL) {
1881 if (db->db_blkid == DMU_BONUS_BLKID) {
1882 dbuf_fix_old_data(db, tx->tx_txg);
1883 data_old = db->db.db_data;
1884 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1886 * Release the data buffer from the cache so
1887 * that we can modify it without impacting
1888 * possible other users of this cached data
1889 * block. Note that indirect blocks and
1890 * private objects are not released until the
1891 * syncing state (since they are only modified
1894 arc_release(db->db_buf, db);
1895 dbuf_fix_old_data(db, tx->tx_txg);
1896 data_old = db->db_buf;
1898 ASSERT(data_old != NULL);
1900 dr->dt.dl.dr_data = data_old;
1902 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1903 list_create(&dr->dt.di.dr_children,
1904 sizeof (dbuf_dirty_record_t),
1905 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1907 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1908 dr->dr_accounted = db->db.db_size;
1910 dr->dr_txg = tx->tx_txg;
1915 * We could have been freed_in_flight between the dbuf_noread
1916 * and dbuf_dirty. We win, as though the dbuf_noread() had
1917 * happened after the free.
1919 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1920 db->db_blkid != DMU_SPILL_BLKID) {
1921 mutex_enter(&dn->dn_mtx);
1922 if (dn->dn_free_ranges[txgoff] != NULL) {
1923 range_tree_clear(dn->dn_free_ranges[txgoff],
1926 mutex_exit(&dn->dn_mtx);
1927 db->db_freed_in_flight = FALSE;
1931 * This buffer is now part of this txg
1933 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1934 db->db_dirtycnt += 1;
1935 ASSERT3U(db->db_dirtycnt, <=, 3);
1937 mutex_exit(&db->db_mtx);
1939 if (db->db_blkid == DMU_BONUS_BLKID ||
1940 db->db_blkid == DMU_SPILL_BLKID) {
1941 mutex_enter(&dn->dn_mtx);
1942 ASSERT(!list_link_active(&dr->dr_dirty_node));
1943 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1944 mutex_exit(&dn->dn_mtx);
1945 dnode_setdirty(dn, tx);
1951 * The dn_struct_rwlock prevents db_blkptr from changing
1952 * due to a write from syncing context completing
1953 * while we are running, so we want to acquire it before
1954 * looking at db_blkptr.
1956 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1957 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1958 drop_struct_lock = TRUE;
1962 * We need to hold the dn_struct_rwlock to make this assertion,
1963 * because it protects dn_phys / dn_next_nlevels from changing.
1965 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1966 dn->dn_phys->dn_nlevels > db->db_level ||
1967 dn->dn_next_nlevels[txgoff] > db->db_level ||
1968 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1969 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1972 * If we are overwriting a dedup BP, then unless it is snapshotted,
1973 * when we get to syncing context we will need to decrement its
1974 * refcount in the DDT. Prefetch the relevant DDT block so that
1975 * syncing context won't have to wait for the i/o.
1977 ddt_prefetch(os->os_spa, db->db_blkptr);
1979 if (db->db_level == 0) {
1980 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1981 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1984 if (db->db_level+1 < dn->dn_nlevels) {
1985 dmu_buf_impl_t *parent = db->db_parent;
1986 dbuf_dirty_record_t *di;
1987 int parent_held = FALSE;
1989 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1990 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1992 parent = dbuf_hold_level(dn, db->db_level+1,
1993 db->db_blkid >> epbs, FTAG);
1994 ASSERT(parent != NULL);
1997 if (drop_struct_lock)
1998 rw_exit(&dn->dn_struct_rwlock);
1999 ASSERT3U(db->db_level+1, ==, parent->db_level);
2000 di = dbuf_dirty(parent, tx);
2002 dbuf_rele(parent, FTAG);
2004 mutex_enter(&db->db_mtx);
2006 * Since we've dropped the mutex, it's possible that
2007 * dbuf_undirty() might have changed this out from under us.
2009 if (db->db_last_dirty == dr ||
2010 dn->dn_object == DMU_META_DNODE_OBJECT) {
2011 mutex_enter(&di->dt.di.dr_mtx);
2012 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
2013 ASSERT(!list_link_active(&dr->dr_dirty_node));
2014 list_insert_tail(&di->dt.di.dr_children, dr);
2015 mutex_exit(&di->dt.di.dr_mtx);
2018 mutex_exit(&db->db_mtx);
2020 ASSERT(db->db_level+1 == dn->dn_nlevels);
2021 ASSERT(db->db_blkid < dn->dn_nblkptr);
2022 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
2023 mutex_enter(&dn->dn_mtx);
2024 ASSERT(!list_link_active(&dr->dr_dirty_node));
2025 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
2026 mutex_exit(&dn->dn_mtx);
2027 if (drop_struct_lock)
2028 rw_exit(&dn->dn_struct_rwlock);
2031 dnode_setdirty(dn, tx);
2037 * Undirty a buffer in the transaction group referenced by the given
2038 * transaction. Return whether this evicted the dbuf.
2041 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
2044 uint64_t txg = tx->tx_txg;
2045 dbuf_dirty_record_t *dr, **drp;
2050 * Due to our use of dn_nlevels below, this can only be called
2051 * in open context, unless we are operating on the MOS.
2052 * From syncing context, dn_nlevels may be different from the
2053 * dn_nlevels used when dbuf was dirtied.
2055 ASSERT(db->db_objset ==
2056 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
2057 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
2058 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2059 ASSERT0(db->db_level);
2060 ASSERT(MUTEX_HELD(&db->db_mtx));
2063 * If this buffer is not dirty, we're done.
2065 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
2066 if (dr->dr_txg <= txg)
2068 if (dr == NULL || dr->dr_txg < txg)
2070 ASSERT(dr->dr_txg == txg);
2071 ASSERT(dr->dr_dbuf == db);
2076 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
2078 ASSERT(db->db.db_size != 0);
2080 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
2081 dr->dr_accounted, txg);
2086 * Note that there are three places in dbuf_dirty()
2087 * where this dirty record may be put on a list.
2088 * Make sure to do a list_remove corresponding to
2089 * every one of those list_insert calls.
2091 if (dr->dr_parent) {
2092 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
2093 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
2094 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
2095 } else if (db->db_blkid == DMU_SPILL_BLKID ||
2096 db->db_level + 1 == dn->dn_nlevels) {
2097 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
2098 mutex_enter(&dn->dn_mtx);
2099 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
2100 mutex_exit(&dn->dn_mtx);
2104 if (db->db_state != DB_NOFILL) {
2105 dbuf_unoverride(dr);
2107 ASSERT(db->db_buf != NULL);
2108 ASSERT(dr->dt.dl.dr_data != NULL);
2109 if (dr->dt.dl.dr_data != db->db_buf)
2110 arc_buf_destroy(dr->dt.dl.dr_data, db);
2113 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2115 ASSERT(db->db_dirtycnt > 0);
2116 db->db_dirtycnt -= 1;
2118 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
2119 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
2128 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
2130 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2131 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
2133 ASSERT(tx->tx_txg != 0);
2134 ASSERT(!refcount_is_zero(&db->db_holds));
2137 * Quick check for dirtyness. For already dirty blocks, this
2138 * reduces runtime of this function by >90%, and overall performance
2139 * by 50% for some workloads (e.g. file deletion with indirect blocks
2142 mutex_enter(&db->db_mtx);
2143 dbuf_dirty_record_t *dr;
2144 for (dr = db->db_last_dirty;
2145 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
2147 * It's possible that it is already dirty but not cached,
2148 * because there are some calls to dbuf_dirty() that don't
2149 * go through dmu_buf_will_dirty().
2151 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
2152 /* This dbuf is already dirty and cached. */
2154 mutex_exit(&db->db_mtx);
2158 mutex_exit(&db->db_mtx);
2161 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
2162 rf |= DB_RF_HAVESTRUCT;
2164 (void) dbuf_read(db, NULL, rf);
2165 (void) dbuf_dirty(db, tx);
2169 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2171 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2173 db->db_state = DB_NOFILL;
2175 dmu_buf_will_fill(db_fake, tx);
2179 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2181 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2183 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2184 ASSERT(tx->tx_txg != 0);
2185 ASSERT(db->db_level == 0);
2186 ASSERT(!refcount_is_zero(&db->db_holds));
2188 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
2189 dmu_tx_private_ok(tx));
2192 (void) dbuf_dirty(db, tx);
2195 #pragma weak dmu_buf_fill_done = dbuf_fill_done
2198 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
2200 mutex_enter(&db->db_mtx);
2203 if (db->db_state == DB_FILL) {
2204 if (db->db_level == 0 && db->db_freed_in_flight) {
2205 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2206 /* we were freed while filling */
2207 /* XXX dbuf_undirty? */
2208 bzero(db->db.db_data, db->db.db_size);
2209 db->db_freed_in_flight = FALSE;
2211 db->db_state = DB_CACHED;
2212 cv_broadcast(&db->db_changed);
2214 mutex_exit(&db->db_mtx);
2218 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
2219 bp_embedded_type_t etype, enum zio_compress comp,
2220 int uncompressed_size, int compressed_size, int byteorder,
2223 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2224 struct dirty_leaf *dl;
2225 dmu_object_type_t type;
2227 if (etype == BP_EMBEDDED_TYPE_DATA) {
2228 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
2229 SPA_FEATURE_EMBEDDED_DATA));
2233 type = DB_DNODE(db)->dn_type;
2236 ASSERT0(db->db_level);
2237 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2239 dmu_buf_will_not_fill(dbuf, tx);
2241 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
2242 dl = &db->db_last_dirty->dt.dl;
2243 encode_embedded_bp_compressed(&dl->dr_overridden_by,
2244 data, comp, uncompressed_size, compressed_size);
2245 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
2246 BP_SET_TYPE(&dl->dr_overridden_by, type);
2247 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
2248 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
2250 dl->dr_override_state = DR_OVERRIDDEN;
2251 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
2255 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2256 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2259 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
2261 ASSERT(!refcount_is_zero(&db->db_holds));
2262 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2263 ASSERT(db->db_level == 0);
2264 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
2265 ASSERT(buf != NULL);
2266 ASSERT(arc_buf_lsize(buf) == db->db.db_size);
2267 ASSERT(tx->tx_txg != 0);
2269 arc_return_buf(buf, db);
2270 ASSERT(arc_released(buf));
2272 mutex_enter(&db->db_mtx);
2274 while (db->db_state == DB_READ || db->db_state == DB_FILL)
2275 cv_wait(&db->db_changed, &db->db_mtx);
2277 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
2279 if (db->db_state == DB_CACHED &&
2280 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
2281 mutex_exit(&db->db_mtx);
2282 (void) dbuf_dirty(db, tx);
2283 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
2284 arc_buf_destroy(buf, db);
2285 xuio_stat_wbuf_copied();
2289 xuio_stat_wbuf_nocopy();
2290 if (db->db_state == DB_CACHED) {
2291 dbuf_dirty_record_t *dr = db->db_last_dirty;
2293 ASSERT(db->db_buf != NULL);
2294 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2295 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2296 if (!arc_released(db->db_buf)) {
2297 ASSERT(dr->dt.dl.dr_override_state ==
2299 arc_release(db->db_buf, db);
2301 dr->dt.dl.dr_data = buf;
2302 arc_buf_destroy(db->db_buf, db);
2303 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2304 arc_release(db->db_buf, db);
2305 arc_buf_destroy(db->db_buf, db);
2309 ASSERT(db->db_buf == NULL);
2310 dbuf_set_data(db, buf);
2311 db->db_state = DB_FILL;
2312 mutex_exit(&db->db_mtx);
2313 (void) dbuf_dirty(db, tx);
2314 dmu_buf_fill_done(&db->db, tx);
2318 dbuf_destroy(dmu_buf_impl_t *db)
2321 dmu_buf_impl_t *parent = db->db_parent;
2322 dmu_buf_impl_t *dndb;
2324 ASSERT(MUTEX_HELD(&db->db_mtx));
2325 ASSERT(refcount_is_zero(&db->db_holds));
2327 if (db->db_buf != NULL) {
2328 arc_buf_destroy(db->db_buf, db);
2332 if (db->db_blkid == DMU_BONUS_BLKID) {
2333 int slots = DB_DNODE(db)->dn_num_slots;
2334 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
2335 if (db->db.db_data != NULL) {
2336 zio_buf_free(db->db.db_data, bonuslen);
2337 arc_space_return(bonuslen, ARC_SPACE_BONUS);
2338 db->db_state = DB_UNCACHED;
2342 dbuf_clear_data(db);
2344 if (multilist_link_active(&db->db_cache_link)) {
2345 ASSERT(db->db_caching_status == DB_DBUF_CACHE ||
2346 db->db_caching_status == DB_DBUF_METADATA_CACHE);
2348 multilist_remove(dbuf_caches[db->db_caching_status].cache, db);
2349 (void) refcount_remove_many(
2350 &dbuf_caches[db->db_caching_status].size,
2351 db->db.db_size, db);
2353 if (db->db_caching_status == DB_DBUF_METADATA_CACHE) {
2354 DBUF_STAT_BUMPDOWN(metadata_cache_count);
2356 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
2357 DBUF_STAT_BUMPDOWN(cache_count);
2358 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
2361 db->db_caching_status = DB_NO_CACHE;
2364 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2365 ASSERT(db->db_data_pending == NULL);
2367 db->db_state = DB_EVICTING;
2368 db->db_blkptr = NULL;
2371 * Now that db_state is DB_EVICTING, nobody else can find this via
2372 * the hash table. We can now drop db_mtx, which allows us to
2373 * acquire the dn_dbufs_mtx.
2375 mutex_exit(&db->db_mtx);
2380 if (db->db_blkid != DMU_BONUS_BLKID) {
2381 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2383 mutex_enter(&dn->dn_dbufs_mtx);
2384 avl_remove(&dn->dn_dbufs, db);
2385 atomic_dec_32(&dn->dn_dbufs_count);
2389 mutex_exit(&dn->dn_dbufs_mtx);
2391 * Decrementing the dbuf count means that the hold corresponding
2392 * to the removed dbuf is no longer discounted in dnode_move(),
2393 * so the dnode cannot be moved until after we release the hold.
2394 * The membar_producer() ensures visibility of the decremented
2395 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2398 mutex_enter(&dn->dn_mtx);
2399 dnode_rele_and_unlock(dn, db, B_TRUE);
2400 db->db_dnode_handle = NULL;
2402 dbuf_hash_remove(db);
2407 ASSERT(refcount_is_zero(&db->db_holds));
2409 db->db_parent = NULL;
2411 ASSERT(db->db_buf == NULL);
2412 ASSERT(db->db.db_data == NULL);
2413 ASSERT(db->db_hash_next == NULL);
2414 ASSERT(db->db_blkptr == NULL);
2415 ASSERT(db->db_data_pending == NULL);
2416 ASSERT3U(db->db_caching_status, ==, DB_NO_CACHE);
2417 ASSERT(!multilist_link_active(&db->db_cache_link));
2419 kmem_cache_free(dbuf_kmem_cache, db);
2420 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2423 * If this dbuf is referenced from an indirect dbuf,
2424 * decrement the ref count on the indirect dbuf.
2426 if (parent && parent != dndb) {
2427 mutex_enter(&parent->db_mtx);
2428 dbuf_rele_and_unlock(parent, db, B_TRUE);
2433 * Note: While bpp will always be updated if the function returns success,
2434 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2435 * this happens when the dnode is the meta-dnode, or a userused or groupused
2438 __attribute__((always_inline))
2440 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2441 dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh)
2446 ASSERT(blkid != DMU_BONUS_BLKID);
2448 if (blkid == DMU_SPILL_BLKID) {
2449 mutex_enter(&dn->dn_mtx);
2450 if (dn->dn_have_spill &&
2451 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2452 *bpp = DN_SPILL_BLKPTR(dn->dn_phys);
2455 dbuf_add_ref(dn->dn_dbuf, NULL);
2456 *parentp = dn->dn_dbuf;
2457 mutex_exit(&dn->dn_mtx);
2462 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2463 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2465 ASSERT3U(level * epbs, <, 64);
2466 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2468 * This assertion shouldn't trip as long as the max indirect block size
2469 * is less than 1M. The reason for this is that up to that point,
2470 * the number of levels required to address an entire object with blocks
2471 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2472 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2473 * (i.e. we can address the entire object), objects will all use at most
2474 * N-1 levels and the assertion won't overflow. However, once epbs is
2475 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2476 * enough to address an entire object, so objects will have 5 levels,
2477 * but then this assertion will overflow.
2479 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2480 * need to redo this logic to handle overflows.
2482 ASSERT(level >= nlevels ||
2483 ((nlevels - level - 1) * epbs) +
2484 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2485 if (level >= nlevels ||
2486 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2487 ((nlevels - level - 1) * epbs)) ||
2489 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2490 /* the buffer has no parent yet */
2491 return (SET_ERROR(ENOENT));
2492 } else if (level < nlevels-1) {
2493 /* this block is referenced from an indirect block */
2496 err = dbuf_hold_impl(dn, level+1,
2497 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2499 __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1,
2500 blkid >> epbs, fail_sparse, FALSE, NULL,
2501 parentp, dh->dh_depth + 1);
2502 err = __dbuf_hold_impl(dh + 1);
2506 err = dbuf_read(*parentp, NULL,
2507 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2509 dbuf_rele(*parentp, NULL);
2513 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2514 (blkid & ((1ULL << epbs) - 1));
2515 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2516 ASSERT(BP_IS_HOLE(*bpp));
2519 /* the block is referenced from the dnode */
2520 ASSERT3U(level, ==, nlevels-1);
2521 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2522 blkid < dn->dn_phys->dn_nblkptr);
2524 dbuf_add_ref(dn->dn_dbuf, NULL);
2525 *parentp = dn->dn_dbuf;
2527 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2532 static dmu_buf_impl_t *
2533 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2534 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2536 objset_t *os = dn->dn_objset;
2537 dmu_buf_impl_t *db, *odb;
2539 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2540 ASSERT(dn->dn_type != DMU_OT_NONE);
2542 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2545 db->db.db_object = dn->dn_object;
2546 db->db_level = level;
2547 db->db_blkid = blkid;
2548 db->db_last_dirty = NULL;
2549 db->db_dirtycnt = 0;
2550 db->db_dnode_handle = dn->dn_handle;
2551 db->db_parent = parent;
2552 db->db_blkptr = blkptr;
2555 db->db_user_immediate_evict = FALSE;
2556 db->db_freed_in_flight = FALSE;
2557 db->db_pending_evict = FALSE;
2559 if (blkid == DMU_BONUS_BLKID) {
2560 ASSERT3P(parent, ==, dn->dn_dbuf);
2561 db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
2562 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2563 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2564 db->db.db_offset = DMU_BONUS_BLKID;
2565 db->db_state = DB_UNCACHED;
2566 db->db_caching_status = DB_NO_CACHE;
2567 /* the bonus dbuf is not placed in the hash table */
2568 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2570 } else if (blkid == DMU_SPILL_BLKID) {
2571 db->db.db_size = (blkptr != NULL) ?
2572 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2573 db->db.db_offset = 0;
2576 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2577 db->db.db_size = blocksize;
2578 db->db.db_offset = db->db_blkid * blocksize;
2582 * Hold the dn_dbufs_mtx while we get the new dbuf
2583 * in the hash table *and* added to the dbufs list.
2584 * This prevents a possible deadlock with someone
2585 * trying to look up this dbuf before its added to the
2588 mutex_enter(&dn->dn_dbufs_mtx);
2589 db->db_state = DB_EVICTING;
2590 if ((odb = dbuf_hash_insert(db)) != NULL) {
2591 /* someone else inserted it first */
2592 kmem_cache_free(dbuf_kmem_cache, db);
2593 mutex_exit(&dn->dn_dbufs_mtx);
2594 DBUF_STAT_BUMP(hash_insert_race);
2597 avl_add(&dn->dn_dbufs, db);
2599 db->db_state = DB_UNCACHED;
2600 db->db_caching_status = DB_NO_CACHE;
2601 mutex_exit(&dn->dn_dbufs_mtx);
2602 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2604 if (parent && parent != dn->dn_dbuf)
2605 dbuf_add_ref(parent, db);
2607 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2608 refcount_count(&dn->dn_holds) > 0);
2609 (void) refcount_add(&dn->dn_holds, db);
2610 atomic_inc_32(&dn->dn_dbufs_count);
2612 dprintf_dbuf(db, "db=%p\n", db);
2617 typedef struct dbuf_prefetch_arg {
2618 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2619 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2620 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2621 int dpa_curlevel; /* The current level that we're reading */
2622 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2623 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2624 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2625 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2626 } dbuf_prefetch_arg_t;
2629 * Actually issue the prefetch read for the block given.
2632 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2634 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2637 arc_flags_t aflags =
2638 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2640 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2641 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2642 ASSERT(dpa->dpa_zio != NULL);
2643 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2644 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2645 &aflags, &dpa->dpa_zb);
2649 * Called when an indirect block above our prefetch target is read in. This
2650 * will either read in the next indirect block down the tree or issue the actual
2651 * prefetch if the next block down is our target.
2654 dbuf_prefetch_indirect_done(zio_t *zio, const zbookmark_phys_t *zb,
2655 const blkptr_t *iobp, arc_buf_t *abuf, void *private)
2657 dbuf_prefetch_arg_t *dpa = private;
2659 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2660 ASSERT3S(dpa->dpa_curlevel, >, 0);
2663 ASSERT(zio == NULL || zio->io_error != 0);
2664 kmem_free(dpa, sizeof (*dpa));
2667 ASSERT(zio == NULL || zio->io_error == 0);
2670 * The dpa_dnode is only valid if we are called with a NULL
2671 * zio. This indicates that the arc_read() returned without
2672 * first calling zio_read() to issue a physical read. Once
2673 * a physical read is made the dpa_dnode must be invalidated
2674 * as the locks guarding it may have been dropped. If the
2675 * dpa_dnode is still valid, then we want to add it to the dbuf
2676 * cache. To do so, we must hold the dbuf associated with the block
2677 * we just prefetched, read its contents so that we associate it
2678 * with an arc_buf_t, and then release it.
2681 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2682 if (zio->io_flags & ZIO_FLAG_RAW) {
2683 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2685 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2687 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2689 dpa->dpa_dnode = NULL;
2690 } else if (dpa->dpa_dnode != NULL) {
2691 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2692 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2693 dpa->dpa_zb.zb_level));
2694 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2695 dpa->dpa_curlevel, curblkid, FTAG);
2696 (void) dbuf_read(db, NULL,
2697 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2698 dbuf_rele(db, FTAG);
2702 kmem_free(dpa, sizeof(*dpa));
2706 dpa->dpa_curlevel--;
2708 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2709 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2710 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2711 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2712 if (BP_IS_HOLE(bp)) {
2713 kmem_free(dpa, sizeof (*dpa));
2714 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2715 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2716 dbuf_issue_final_prefetch(dpa, bp);
2717 kmem_free(dpa, sizeof (*dpa));
2719 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2720 zbookmark_phys_t zb;
2722 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2723 if (dpa->dpa_aflags & ARC_FLAG_L2CACHE)
2724 iter_aflags |= ARC_FLAG_L2CACHE;
2726 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2728 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2729 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2731 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2732 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2733 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2737 arc_buf_destroy(abuf, private);
2741 * Issue prefetch reads for the given block on the given level. If the indirect
2742 * blocks above that block are not in memory, we will read them in
2743 * asynchronously. As a result, this call never blocks waiting for a read to
2747 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2751 int epbs, nlevels, curlevel;
2754 ASSERT(blkid != DMU_BONUS_BLKID);
2755 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2757 if (blkid > dn->dn_maxblkid)
2760 if (dnode_block_freed(dn, blkid))
2764 * This dnode hasn't been written to disk yet, so there's nothing to
2767 nlevels = dn->dn_phys->dn_nlevels;
2768 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2771 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2772 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2775 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2778 mutex_exit(&db->db_mtx);
2780 * This dbuf already exists. It is either CACHED, or
2781 * (we assume) about to be read or filled.
2787 * Find the closest ancestor (indirect block) of the target block
2788 * that is present in the cache. In this indirect block, we will
2789 * find the bp that is at curlevel, curblkid.
2793 while (curlevel < nlevels - 1) {
2794 int parent_level = curlevel + 1;
2795 uint64_t parent_blkid = curblkid >> epbs;
2798 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2799 FALSE, TRUE, FTAG, &db) == 0) {
2800 blkptr_t *bpp = db->db_buf->b_data;
2801 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2802 dbuf_rele(db, FTAG);
2806 curlevel = parent_level;
2807 curblkid = parent_blkid;
2810 if (curlevel == nlevels - 1) {
2811 /* No cached indirect blocks found. */
2812 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2813 bp = dn->dn_phys->dn_blkptr[curblkid];
2815 if (BP_IS_HOLE(&bp))
2818 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2820 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2823 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2824 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2825 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2826 dn->dn_object, level, blkid);
2827 dpa->dpa_curlevel = curlevel;
2828 dpa->dpa_prio = prio;
2829 dpa->dpa_aflags = aflags;
2830 dpa->dpa_spa = dn->dn_objset->os_spa;
2831 dpa->dpa_dnode = dn;
2832 dpa->dpa_epbs = epbs;
2835 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2836 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2837 dpa->dpa_aflags |= ARC_FLAG_L2CACHE;
2840 * If we have the indirect just above us, no need to do the asynchronous
2841 * prefetch chain; we'll just run the last step ourselves. If we're at
2842 * a higher level, though, we want to issue the prefetches for all the
2843 * indirect blocks asynchronously, so we can go on with whatever we were
2846 if (curlevel == level) {
2847 ASSERT3U(curblkid, ==, blkid);
2848 dbuf_issue_final_prefetch(dpa, &bp);
2849 kmem_free(dpa, sizeof (*dpa));
2851 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2852 zbookmark_phys_t zb;
2854 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2855 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2856 iter_aflags |= ARC_FLAG_L2CACHE;
2858 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2859 dn->dn_object, curlevel, curblkid);
2860 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2861 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2862 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2866 * We use pio here instead of dpa_zio since it's possible that
2867 * dpa may have already been freed.
2872 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2875 * Helper function for __dbuf_hold_impl() to copy a buffer. Handles
2876 * the case of encrypted, compressed and uncompressed buffers by
2877 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
2878 * arc_alloc_compressed_buf() or arc_alloc_buf().*
2880 * NOTE: Declared noinline to avoid stack bloat in __dbuf_hold_impl().
2882 noinline static void
2883 dbuf_hold_copy(struct dbuf_hold_impl_data *dh)
2885 dnode_t *dn = dh->dh_dn;
2886 dmu_buf_impl_t *db = dh->dh_db;
2887 dbuf_dirty_record_t *dr = dh->dh_dr;
2888 arc_buf_t *data = dr->dt.dl.dr_data;
2890 enum zio_compress compress_type = arc_get_compression(data);
2892 if (compress_type != ZIO_COMPRESS_OFF) {
2893 dbuf_set_data(db, arc_alloc_compressed_buf(
2894 dn->dn_objset->os_spa, db, arc_buf_size(data),
2895 arc_buf_lsize(data), compress_type));
2897 dbuf_set_data(db, arc_alloc_buf(dn->dn_objset->os_spa, db,
2898 DBUF_GET_BUFC_TYPE(db), db->db.db_size));
2901 bcopy(data->b_data, db->db.db_data, arc_buf_size(data));
2905 * Returns with db_holds incremented, and db_mtx not held.
2906 * Note: dn_struct_rwlock must be held.
2909 __dbuf_hold_impl(struct dbuf_hold_impl_data *dh)
2911 ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH);
2912 dh->dh_parent = NULL;
2914 ASSERT(dh->dh_blkid != DMU_BONUS_BLKID);
2915 ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock));
2916 ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level);
2918 *(dh->dh_dbp) = NULL;
2920 /* dbuf_find() returns with db_mtx held */
2921 dh->dh_db = dbuf_find(dh->dh_dn->dn_objset, dh->dh_dn->dn_object,
2922 dh->dh_level, dh->dh_blkid);
2924 if (dh->dh_db == NULL) {
2927 if (dh->dh_fail_uncached)
2928 return (SET_ERROR(ENOENT));
2930 ASSERT3P(dh->dh_parent, ==, NULL);
2931 dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2932 dh->dh_fail_sparse, &dh->dh_parent, &dh->dh_bp, dh);
2933 if (dh->dh_fail_sparse) {
2934 if (dh->dh_err == 0 &&
2935 dh->dh_bp && BP_IS_HOLE(dh->dh_bp))
2936 dh->dh_err = SET_ERROR(ENOENT);
2939 dbuf_rele(dh->dh_parent, NULL);
2940 return (dh->dh_err);
2943 if (dh->dh_err && dh->dh_err != ENOENT)
2944 return (dh->dh_err);
2945 dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2946 dh->dh_parent, dh->dh_bp);
2949 if (dh->dh_fail_uncached && dh->dh_db->db_state != DB_CACHED) {
2950 mutex_exit(&dh->dh_db->db_mtx);
2951 return (SET_ERROR(ENOENT));
2954 if (dh->dh_db->db_buf != NULL) {
2955 arc_buf_access(dh->dh_db->db_buf);
2956 ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data);
2959 ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf));
2962 * If this buffer is currently syncing out, and we are are
2963 * still referencing it from db_data, we need to make a copy
2964 * of it in case we decide we want to dirty it again in this txg.
2966 if (dh->dh_db->db_level == 0 &&
2967 dh->dh_db->db_blkid != DMU_BONUS_BLKID &&
2968 dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT &&
2969 dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) {
2970 dh->dh_dr = dh->dh_db->db_data_pending;
2971 if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf)
2975 if (multilist_link_active(&dh->dh_db->db_cache_link)) {
2976 ASSERT(refcount_is_zero(&dh->dh_db->db_holds));
2977 ASSERT(dh->dh_db->db_caching_status == DB_DBUF_CACHE ||
2978 dh->dh_db->db_caching_status == DB_DBUF_METADATA_CACHE);
2981 dbuf_caches[dh->dh_db->db_caching_status].cache,
2983 (void) refcount_remove_many(
2984 &dbuf_caches[dh->dh_db->db_caching_status].size,
2985 dh->dh_db->db.db_size, dh->dh_db);
2987 if (dh->dh_db->db_caching_status == DB_DBUF_METADATA_CACHE) {
2988 DBUF_STAT_BUMPDOWN(metadata_cache_count);
2990 DBUF_STAT_BUMPDOWN(cache_levels[dh->dh_db->db_level]);
2991 DBUF_STAT_BUMPDOWN(cache_count);
2992 DBUF_STAT_DECR(cache_levels_bytes[dh->dh_db->db_level],
2993 dh->dh_db->db.db_size);
2995 dh->dh_db->db_caching_status = DB_NO_CACHE;
2997 (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag);
2998 DBUF_VERIFY(dh->dh_db);
2999 mutex_exit(&dh->dh_db->db_mtx);
3001 /* NOTE: we can't rele the parent until after we drop the db_mtx */
3003 dbuf_rele(dh->dh_parent, NULL);
3005 ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn);
3006 ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid);
3007 ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level);
3008 *(dh->dh_dbp) = dh->dh_db;
3014 * The following code preserves the recursive function dbuf_hold_impl()
3015 * but moves the local variables AND function arguments to the heap to
3016 * minimize the stack frame size. Enough space is initially allocated
3017 * on the stack for 20 levels of recursion.
3020 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
3021 boolean_t fail_sparse, boolean_t fail_uncached,
3022 void *tag, dmu_buf_impl_t **dbp)
3024 struct dbuf_hold_impl_data *dh;
3027 dh = kmem_alloc(sizeof (struct dbuf_hold_impl_data) *
3028 DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP);
3029 __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse,
3030 fail_uncached, tag, dbp, 0);
3032 error = __dbuf_hold_impl(dh);
3034 kmem_free(dh, sizeof (struct dbuf_hold_impl_data) *
3035 DBUF_HOLD_IMPL_MAX_DEPTH);
3041 __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
3042 dnode_t *dn, uint8_t level, uint64_t blkid,
3043 boolean_t fail_sparse, boolean_t fail_uncached,
3044 void *tag, dmu_buf_impl_t **dbp, int depth)
3047 dh->dh_level = level;
3048 dh->dh_blkid = blkid;
3050 dh->dh_fail_sparse = fail_sparse;
3051 dh->dh_fail_uncached = fail_uncached;
3057 dh->dh_parent = NULL;
3062 dh->dh_depth = depth;
3066 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
3068 return (dbuf_hold_level(dn, 0, blkid, tag));
3072 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
3075 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
3076 return (err ? NULL : db);
3080 dbuf_create_bonus(dnode_t *dn)
3082 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
3084 ASSERT(dn->dn_bonus == NULL);
3085 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
3089 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
3091 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3094 if (db->db_blkid != DMU_SPILL_BLKID)
3095 return (SET_ERROR(ENOTSUP));
3097 blksz = SPA_MINBLOCKSIZE;
3098 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
3099 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
3103 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3104 dbuf_new_size(db, blksz, tx);
3105 rw_exit(&dn->dn_struct_rwlock);
3112 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
3114 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
3117 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3119 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
3121 int64_t holds = refcount_add(&db->db_holds, tag);
3122 ASSERT3S(holds, >, 1);
3125 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3127 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
3130 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3131 dmu_buf_impl_t *found_db;
3132 boolean_t result = B_FALSE;
3134 if (db->db_blkid == DMU_BONUS_BLKID)
3135 found_db = dbuf_find_bonus(os, obj);
3137 found_db = dbuf_find(os, obj, 0, blkid);
3139 if (found_db != NULL) {
3140 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
3141 (void) refcount_add(&db->db_holds, tag);
3144 mutex_exit(&db->db_mtx);
3150 * If you call dbuf_rele() you had better not be referencing the dnode handle
3151 * unless you have some other direct or indirect hold on the dnode. (An indirect
3152 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3153 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3154 * dnode's parent dbuf evicting its dnode handles.
3157 dbuf_rele(dmu_buf_impl_t *db, void *tag)
3159 mutex_enter(&db->db_mtx);
3160 dbuf_rele_and_unlock(db, tag, B_FALSE);
3164 dmu_buf_rele(dmu_buf_t *db, void *tag)
3166 dbuf_rele((dmu_buf_impl_t *)db, tag);
3170 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3171 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
3172 * argument should be set if we are already in the dbuf-evicting code
3173 * path, in which case we don't want to recursively evict. This allows us to
3174 * avoid deeply nested stacks that would have a call flow similar to this:
3176 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
3179 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
3183 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag, boolean_t evicting)
3187 ASSERT(MUTEX_HELD(&db->db_mtx));
3191 * Remove the reference to the dbuf before removing its hold on the
3192 * dnode so we can guarantee in dnode_move() that a referenced bonus
3193 * buffer has a corresponding dnode hold.
3195 holds = refcount_remove(&db->db_holds, tag);
3199 * We can't freeze indirects if there is a possibility that they
3200 * may be modified in the current syncing context.
3202 if (db->db_buf != NULL &&
3203 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
3204 arc_buf_freeze(db->db_buf);
3207 if (holds == db->db_dirtycnt &&
3208 db->db_level == 0 && db->db_user_immediate_evict)
3209 dbuf_evict_user(db);
3212 if (db->db_blkid == DMU_BONUS_BLKID) {
3214 boolean_t evict_dbuf = db->db_pending_evict;
3217 * If the dnode moves here, we cannot cross this
3218 * barrier until the move completes.
3223 atomic_dec_32(&dn->dn_dbufs_count);
3226 * Decrementing the dbuf count means that the bonus
3227 * buffer's dnode hold is no longer discounted in
3228 * dnode_move(). The dnode cannot move until after
3229 * the dnode_rele() below.
3234 * Do not reference db after its lock is dropped.
3235 * Another thread may evict it.
3237 mutex_exit(&db->db_mtx);
3240 dnode_evict_bonus(dn);
3243 } else if (db->db_buf == NULL) {
3245 * This is a special case: we never associated this
3246 * dbuf with any data allocated from the ARC.
3248 ASSERT(db->db_state == DB_UNCACHED ||
3249 db->db_state == DB_NOFILL);
3251 } else if (arc_released(db->db_buf)) {
3253 * This dbuf has anonymous data associated with it.
3257 boolean_t do_arc_evict = B_FALSE;
3259 spa_t *spa = dmu_objset_spa(db->db_objset);
3261 if (!DBUF_IS_CACHEABLE(db) &&
3262 db->db_blkptr != NULL &&
3263 !BP_IS_HOLE(db->db_blkptr) &&
3264 !BP_IS_EMBEDDED(db->db_blkptr)) {
3265 do_arc_evict = B_TRUE;
3266 bp = *db->db_blkptr;
3269 if (!DBUF_IS_CACHEABLE(db) ||
3270 db->db_pending_evict) {
3272 } else if (!multilist_link_active(&db->db_cache_link)) {
3273 ASSERT3U(db->db_caching_status, ==,
3276 dbuf_cached_state_t dcs =
3277 dbuf_include_in_metadata_cache(db) ?
3278 DB_DBUF_METADATA_CACHE : DB_DBUF_CACHE;
3279 db->db_caching_status = dcs;
3281 multilist_insert(dbuf_caches[dcs].cache, db);
3282 (void) refcount_add_many(&dbuf_caches[dcs].size,
3283 db->db.db_size, db);
3285 if (dcs == DB_DBUF_METADATA_CACHE) {
3286 DBUF_STAT_BUMP(metadata_cache_count);
3288 metadata_cache_size_bytes_max,
3290 &dbuf_caches[dcs].size));
3293 cache_levels[db->db_level]);
3294 DBUF_STAT_BUMP(cache_count);
3296 cache_levels_bytes[db->db_level],
3298 DBUF_STAT_MAX(cache_size_bytes_max,
3300 &dbuf_caches[dcs].size));
3302 mutex_exit(&db->db_mtx);
3304 if (db->db_caching_status == DB_DBUF_CACHE &&
3306 dbuf_evict_notify();
3311 arc_freed(spa, &bp);
3314 mutex_exit(&db->db_mtx);
3319 #pragma weak dmu_buf_refcount = dbuf_refcount
3321 dbuf_refcount(dmu_buf_impl_t *db)
3323 return (refcount_count(&db->db_holds));
3327 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
3328 dmu_buf_user_t *new_user)
3330 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3332 mutex_enter(&db->db_mtx);
3333 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3334 if (db->db_user == old_user)
3335 db->db_user = new_user;
3337 old_user = db->db_user;
3338 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3339 mutex_exit(&db->db_mtx);
3345 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3347 return (dmu_buf_replace_user(db_fake, NULL, user));
3351 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3353 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3355 db->db_user_immediate_evict = TRUE;
3356 return (dmu_buf_set_user(db_fake, user));
3360 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3362 return (dmu_buf_replace_user(db_fake, user, NULL));
3366 dmu_buf_get_user(dmu_buf_t *db_fake)
3368 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3370 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3371 return (db->db_user);
3375 dmu_buf_user_evict_wait()
3377 taskq_wait(dbu_evict_taskq);
3381 dmu_buf_get_blkptr(dmu_buf_t *db)
3383 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3384 return (dbi->db_blkptr);
3388 dmu_buf_get_objset(dmu_buf_t *db)
3390 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3391 return (dbi->db_objset);
3395 dmu_buf_dnode_enter(dmu_buf_t *db)
3397 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3398 DB_DNODE_ENTER(dbi);
3399 return (DB_DNODE(dbi));
3403 dmu_buf_dnode_exit(dmu_buf_t *db)
3405 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3410 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
3412 /* ASSERT(dmu_tx_is_syncing(tx) */
3413 ASSERT(MUTEX_HELD(&db->db_mtx));
3415 if (db->db_blkptr != NULL)
3418 if (db->db_blkid == DMU_SPILL_BLKID) {
3419 db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys);
3420 BP_ZERO(db->db_blkptr);
3423 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
3425 * This buffer was allocated at a time when there was
3426 * no available blkptrs from the dnode, or it was
3427 * inappropriate to hook it in (i.e., nlevels mis-match).
3429 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
3430 ASSERT(db->db_parent == NULL);
3431 db->db_parent = dn->dn_dbuf;
3432 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
3435 dmu_buf_impl_t *parent = db->db_parent;
3436 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3438 ASSERT(dn->dn_phys->dn_nlevels > 1);
3439 if (parent == NULL) {
3440 mutex_exit(&db->db_mtx);
3441 rw_enter(&dn->dn_struct_rwlock, RW_READER);
3442 parent = dbuf_hold_level(dn, db->db_level + 1,
3443 db->db_blkid >> epbs, db);
3444 rw_exit(&dn->dn_struct_rwlock);
3445 mutex_enter(&db->db_mtx);
3446 db->db_parent = parent;
3448 db->db_blkptr = (blkptr_t *)parent->db.db_data +
3449 (db->db_blkid & ((1ULL << epbs) - 1));
3455 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3456 * is critical the we not allow the compiler to inline this function in to
3457 * dbuf_sync_list() thereby drastically bloating the stack usage.
3459 noinline static void
3460 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3462 dmu_buf_impl_t *db = dr->dr_dbuf;
3466 ASSERT(dmu_tx_is_syncing(tx));
3468 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3470 mutex_enter(&db->db_mtx);
3472 ASSERT(db->db_level > 0);
3475 /* Read the block if it hasn't been read yet. */
3476 if (db->db_buf == NULL) {
3477 mutex_exit(&db->db_mtx);
3478 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
3479 mutex_enter(&db->db_mtx);
3481 ASSERT3U(db->db_state, ==, DB_CACHED);
3482 ASSERT(db->db_buf != NULL);
3486 /* Indirect block size must match what the dnode thinks it is. */
3487 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3488 dbuf_check_blkptr(dn, db);
3491 /* Provide the pending dirty record to child dbufs */
3492 db->db_data_pending = dr;
3494 mutex_exit(&db->db_mtx);
3496 dbuf_write(dr, db->db_buf, tx);
3499 mutex_enter(&dr->dt.di.dr_mtx);
3500 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3501 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3502 mutex_exit(&dr->dt.di.dr_mtx);
3507 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3508 * critical the we not allow the compiler to inline this function in to
3509 * dbuf_sync_list() thereby drastically bloating the stack usage.
3511 noinline static void
3512 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3514 arc_buf_t **datap = &dr->dt.dl.dr_data;
3515 dmu_buf_impl_t *db = dr->dr_dbuf;
3518 uint64_t txg = tx->tx_txg;
3520 ASSERT(dmu_tx_is_syncing(tx));
3522 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3524 mutex_enter(&db->db_mtx);
3526 * To be synced, we must be dirtied. But we
3527 * might have been freed after the dirty.
3529 if (db->db_state == DB_UNCACHED) {
3530 /* This buffer has been freed since it was dirtied */
3531 ASSERT(db->db.db_data == NULL);
3532 } else if (db->db_state == DB_FILL) {
3533 /* This buffer was freed and is now being re-filled */
3534 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3536 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3543 if (db->db_blkid == DMU_SPILL_BLKID) {
3544 mutex_enter(&dn->dn_mtx);
3545 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
3547 * In the previous transaction group, the bonus buffer
3548 * was entirely used to store the attributes for the
3549 * dnode which overrode the dn_spill field. However,
3550 * when adding more attributes to the file a spill
3551 * block was required to hold the extra attributes.
3553 * Make sure to clear the garbage left in the dn_spill
3554 * field from the previous attributes in the bonus
3555 * buffer. Otherwise, after writing out the spill
3556 * block to the new allocated dva, it will free
3557 * the old block pointed to by the invalid dn_spill.
3559 db->db_blkptr = NULL;
3561 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3562 mutex_exit(&dn->dn_mtx);
3566 * If this is a bonus buffer, simply copy the bonus data into the
3567 * dnode. It will be written out when the dnode is synced (and it
3568 * will be synced, since it must have been dirty for dbuf_sync to
3571 if (db->db_blkid == DMU_BONUS_BLKID) {
3572 dbuf_dirty_record_t **drp;
3574 ASSERT(*datap != NULL);
3575 ASSERT0(db->db_level);
3576 ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=,
3577 DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1));
3578 bcopy(*datap, DN_BONUS(dn->dn_phys),
3579 DN_MAX_BONUS_LEN(dn->dn_phys));
3582 if (*datap != db->db.db_data) {
3583 int slots = DB_DNODE(db)->dn_num_slots;
3584 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
3585 zio_buf_free(*datap, bonuslen);
3586 arc_space_return(bonuslen, ARC_SPACE_BONUS);
3588 db->db_data_pending = NULL;
3589 drp = &db->db_last_dirty;
3591 drp = &(*drp)->dr_next;
3592 ASSERT(dr->dr_next == NULL);
3593 ASSERT(dr->dr_dbuf == db);
3595 if (dr->dr_dbuf->db_level != 0) {
3596 mutex_destroy(&dr->dt.di.dr_mtx);
3597 list_destroy(&dr->dt.di.dr_children);
3599 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3600 ASSERT(db->db_dirtycnt > 0);
3601 db->db_dirtycnt -= 1;
3602 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE);
3609 * This function may have dropped the db_mtx lock allowing a dmu_sync
3610 * operation to sneak in. As a result, we need to ensure that we
3611 * don't check the dr_override_state until we have returned from
3612 * dbuf_check_blkptr.
3614 dbuf_check_blkptr(dn, db);
3617 * If this buffer is in the middle of an immediate write,
3618 * wait for the synchronous IO to complete.
3620 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3621 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3622 cv_wait(&db->db_changed, &db->db_mtx);
3623 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3626 if (db->db_state != DB_NOFILL &&
3627 dn->dn_object != DMU_META_DNODE_OBJECT &&
3628 refcount_count(&db->db_holds) > 1 &&
3629 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3630 *datap == db->db_buf) {
3632 * If this buffer is currently "in use" (i.e., there
3633 * are active holds and db_data still references it),
3634 * then make a copy before we start the write so that
3635 * any modifications from the open txg will not leak
3638 * NOTE: this copy does not need to be made for
3639 * objects only modified in the syncing context (e.g.
3640 * DNONE_DNODE blocks).
3642 int psize = arc_buf_size(*datap);
3643 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3644 enum zio_compress compress_type = arc_get_compression(*datap);
3646 if (compress_type == ZIO_COMPRESS_OFF) {
3647 *datap = arc_alloc_buf(os->os_spa, db, type, psize);
3649 ASSERT3U(type, ==, ARC_BUFC_DATA);
3650 int lsize = arc_buf_lsize(*datap);
3651 *datap = arc_alloc_compressed_buf(os->os_spa, db,
3652 psize, lsize, compress_type);
3654 bcopy(db->db.db_data, (*datap)->b_data, psize);
3656 db->db_data_pending = dr;
3658 mutex_exit(&db->db_mtx);
3660 dbuf_write(dr, *datap, tx);
3662 ASSERT(!list_link_active(&dr->dr_dirty_node));
3663 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3664 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3668 * Although zio_nowait() does not "wait for an IO", it does
3669 * initiate the IO. If this is an empty write it seems plausible
3670 * that the IO could actually be completed before the nowait
3671 * returns. We need to DB_DNODE_EXIT() first in case
3672 * zio_nowait() invalidates the dbuf.
3675 zio_nowait(dr->dr_zio);
3680 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3682 dbuf_dirty_record_t *dr;
3684 while (dr = list_head(list)) {
3685 if (dr->dr_zio != NULL) {
3687 * If we find an already initialized zio then we
3688 * are processing the meta-dnode, and we have finished.
3689 * The dbufs for all dnodes are put back on the list
3690 * during processing, so that we can zio_wait()
3691 * these IOs after initiating all child IOs.
3693 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3694 DMU_META_DNODE_OBJECT);
3697 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3698 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3699 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3701 list_remove(list, dr);
3702 if (dr->dr_dbuf->db_level > 0)
3703 dbuf_sync_indirect(dr, tx);
3705 dbuf_sync_leaf(dr, tx);
3711 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3713 dmu_buf_impl_t *db = vdb;
3715 blkptr_t *bp = zio->io_bp;
3716 blkptr_t *bp_orig = &zio->io_bp_orig;
3717 spa_t *spa = zio->io_spa;
3722 ASSERT3P(db->db_blkptr, !=, NULL);
3723 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3727 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3728 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3729 zio->io_prev_space_delta = delta;
3731 if (bp->blk_birth != 0) {
3732 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3733 BP_GET_TYPE(bp) == dn->dn_type) ||
3734 (db->db_blkid == DMU_SPILL_BLKID &&
3735 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3736 BP_IS_EMBEDDED(bp));
3737 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3740 mutex_enter(&db->db_mtx);
3743 if (db->db_blkid == DMU_SPILL_BLKID) {
3744 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3745 ASSERT(!(BP_IS_HOLE(bp)) &&
3746 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3750 if (db->db_level == 0) {
3751 mutex_enter(&dn->dn_mtx);
3752 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3753 db->db_blkid != DMU_SPILL_BLKID)
3754 dn->dn_phys->dn_maxblkid = db->db_blkid;
3755 mutex_exit(&dn->dn_mtx);
3757 if (dn->dn_type == DMU_OT_DNODE) {
3759 while (i < db->db.db_size) {
3760 dnode_phys_t *dnp = db->db.db_data + i;
3762 i += DNODE_MIN_SIZE;
3763 if (dnp->dn_type != DMU_OT_NONE) {
3765 i += dnp->dn_extra_slots *
3770 if (BP_IS_HOLE(bp)) {
3777 blkptr_t *ibp = db->db.db_data;
3778 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3779 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3780 if (BP_IS_HOLE(ibp))
3782 fill += BP_GET_FILL(ibp);
3787 if (!BP_IS_EMBEDDED(bp))
3788 bp->blk_fill = fill;
3790 mutex_exit(&db->db_mtx);
3792 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3793 *db->db_blkptr = *bp;
3794 rw_exit(&dn->dn_struct_rwlock);
3799 * This function gets called just prior to running through the compression
3800 * stage of the zio pipeline. If we're an indirect block comprised of only
3801 * holes, then we want this indirect to be compressed away to a hole. In
3802 * order to do that we must zero out any information about the holes that
3803 * this indirect points to prior to before we try to compress it.
3806 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3808 dmu_buf_impl_t *db = vdb;
3811 unsigned int epbs, i;
3813 ASSERT3U(db->db_level, >, 0);
3816 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3817 ASSERT3U(epbs, <, 31);
3819 /* Determine if all our children are holes */
3820 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
3821 if (!BP_IS_HOLE(bp))
3826 * If all the children are holes, then zero them all out so that
3827 * we may get compressed away.
3829 if (i == 1 << epbs) {
3831 * We only found holes. Grab the rwlock to prevent
3832 * anybody from reading the blocks we're about to
3835 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3836 bzero(db->db.db_data, db->db.db_size);
3837 rw_exit(&dn->dn_struct_rwlock);
3843 * The SPA will call this callback several times for each zio - once
3844 * for every physical child i/o (zio->io_phys_children times). This
3845 * allows the DMU to monitor the progress of each logical i/o. For example,
3846 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3847 * block. There may be a long delay before all copies/fragments are completed,
3848 * so this callback allows us to retire dirty space gradually, as the physical
3853 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3855 dmu_buf_impl_t *db = arg;
3856 objset_t *os = db->db_objset;
3857 dsl_pool_t *dp = dmu_objset_pool(os);
3858 dbuf_dirty_record_t *dr;
3861 dr = db->db_data_pending;
3862 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3865 * The callback will be called io_phys_children times. Retire one
3866 * portion of our dirty space each time we are called. Any rounding
3867 * error will be cleaned up by dsl_pool_sync()'s call to
3868 * dsl_pool_undirty_space().
3870 delta = dr->dr_accounted / zio->io_phys_children;
3871 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3876 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3878 dmu_buf_impl_t *db = vdb;
3879 blkptr_t *bp_orig = &zio->io_bp_orig;
3880 blkptr_t *bp = db->db_blkptr;
3881 objset_t *os = db->db_objset;
3882 dmu_tx_t *tx = os->os_synctx;
3883 dbuf_dirty_record_t **drp, *dr;
3885 ASSERT0(zio->io_error);
3886 ASSERT(db->db_blkptr == bp);
3889 * For nopwrites and rewrites we ensure that the bp matches our
3890 * original and bypass all the accounting.
3892 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3893 ASSERT(BP_EQUAL(bp, bp_orig));
3895 dsl_dataset_t *ds = os->os_dsl_dataset;
3896 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3897 dsl_dataset_block_born(ds, bp, tx);
3900 mutex_enter(&db->db_mtx);
3904 drp = &db->db_last_dirty;
3905 while ((dr = *drp) != db->db_data_pending)
3907 ASSERT(!list_link_active(&dr->dr_dirty_node));
3908 ASSERT(dr->dr_dbuf == db);
3909 ASSERT(dr->dr_next == NULL);
3913 if (db->db_blkid == DMU_SPILL_BLKID) {
3918 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3919 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3920 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3925 if (db->db_level == 0) {
3926 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3927 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3928 if (db->db_state != DB_NOFILL) {
3929 if (dr->dt.dl.dr_data != db->db_buf)
3930 arc_buf_destroy(dr->dt.dl.dr_data, db);
3937 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3938 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3939 if (!BP_IS_HOLE(db->db_blkptr)) {
3941 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3942 ASSERT3U(db->db_blkid, <=,
3943 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3944 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3948 mutex_destroy(&dr->dt.di.dr_mtx);
3949 list_destroy(&dr->dt.di.dr_children);
3951 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3953 cv_broadcast(&db->db_changed);
3954 ASSERT(db->db_dirtycnt > 0);
3955 db->db_dirtycnt -= 1;
3956 db->db_data_pending = NULL;
3957 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg, B_FALSE);
3961 dbuf_write_nofill_ready(zio_t *zio)
3963 dbuf_write_ready(zio, NULL, zio->io_private);
3967 dbuf_write_nofill_done(zio_t *zio)
3969 dbuf_write_done(zio, NULL, zio->io_private);
3973 dbuf_write_override_ready(zio_t *zio)
3975 dbuf_dirty_record_t *dr = zio->io_private;
3976 dmu_buf_impl_t *db = dr->dr_dbuf;
3978 dbuf_write_ready(zio, NULL, db);
3982 dbuf_write_override_done(zio_t *zio)
3984 dbuf_dirty_record_t *dr = zio->io_private;
3985 dmu_buf_impl_t *db = dr->dr_dbuf;
3986 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3988 mutex_enter(&db->db_mtx);
3989 if (!BP_EQUAL(zio->io_bp, obp)) {
3990 if (!BP_IS_HOLE(obp))
3991 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3992 arc_release(dr->dt.dl.dr_data, db);
3994 mutex_exit(&db->db_mtx);
3995 dbuf_write_done(zio, NULL, db);
3997 if (zio->io_abd != NULL)
3998 abd_put(zio->io_abd);
4001 typedef struct dbuf_remap_impl_callback_arg {
4003 uint64_t drica_blk_birth;
4005 } dbuf_remap_impl_callback_arg_t;
4008 dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size,
4011 dbuf_remap_impl_callback_arg_t *drica = arg;
4012 objset_t *os = drica->drica_os;
4013 spa_t *spa = dmu_objset_spa(os);
4014 dmu_tx_t *tx = drica->drica_tx;
4016 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4018 if (os == spa_meta_objset(spa)) {
4019 spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx);
4021 dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset,
4022 size, drica->drica_blk_birth, tx);
4027 dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, dmu_tx_t *tx)
4029 blkptr_t bp_copy = *bp;
4030 spa_t *spa = dmu_objset_spa(dn->dn_objset);
4031 dbuf_remap_impl_callback_arg_t drica;
4033 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4035 drica.drica_os = dn->dn_objset;
4036 drica.drica_blk_birth = bp->blk_birth;
4037 drica.drica_tx = tx;
4038 if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback,
4041 * The struct_rwlock prevents dbuf_read_impl() from
4042 * dereferencing the BP while we are changing it. To
4043 * avoid lock contention, only grab it when we are actually
4046 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
4048 rw_exit(&dn->dn_struct_rwlock);
4053 * Returns true if a dbuf_remap would modify the dbuf. We do this by attempting
4054 * to remap a copy of every bp in the dbuf.
4057 dbuf_can_remap(const dmu_buf_impl_t *db)
4059 spa_t *spa = dmu_objset_spa(db->db_objset);
4060 blkptr_t *bp = db->db.db_data;
4061 boolean_t ret = B_FALSE;
4063 ASSERT3U(db->db_level, >, 0);
4064 ASSERT3S(db->db_state, ==, DB_CACHED);
4066 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
4068 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
4069 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
4070 blkptr_t bp_copy = bp[i];
4071 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
4076 spa_config_exit(spa, SCL_VDEV, FTAG);
4082 dnode_needs_remap(const dnode_t *dn)
4084 spa_t *spa = dmu_objset_spa(dn->dn_objset);
4085 boolean_t ret = B_FALSE;
4087 if (dn->dn_phys->dn_nlevels == 0) {
4091 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
4093 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
4094 for (int j = 0; j < dn->dn_phys->dn_nblkptr; j++) {
4095 blkptr_t bp_copy = dn->dn_phys->dn_blkptr[j];
4096 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
4101 spa_config_exit(spa, SCL_VDEV, FTAG);
4107 * Remap any existing BP's to concrete vdevs, if possible.
4110 dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx)
4112 spa_t *spa = dmu_objset_spa(db->db_objset);
4113 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4115 if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL))
4118 if (db->db_level > 0) {
4119 blkptr_t *bp = db->db.db_data;
4120 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
4121 dbuf_remap_impl(dn, &bp[i], tx);
4123 } else if (db->db.db_object == DMU_META_DNODE_OBJECT) {
4124 dnode_phys_t *dnp = db->db.db_data;
4125 ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==,
4127 for (int i = 0; i < db->db.db_size >> DNODE_SHIFT;
4128 i += dnp[i].dn_extra_slots + 1) {
4129 for (int j = 0; j < dnp[i].dn_nblkptr; j++) {
4130 dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], tx);
4137 /* Issue I/O to commit a dirty buffer to disk. */
4139 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
4141 dmu_buf_impl_t *db = dr->dr_dbuf;
4144 dmu_buf_impl_t *parent = db->db_parent;
4145 uint64_t txg = tx->tx_txg;
4146 zbookmark_phys_t zb;
4151 ASSERT(dmu_tx_is_syncing(tx));
4157 if (db->db_state != DB_NOFILL) {
4158 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
4160 * Private object buffers are released here rather
4161 * than in dbuf_dirty() since they are only modified
4162 * in the syncing context and we don't want the
4163 * overhead of making multiple copies of the data.
4165 if (BP_IS_HOLE(db->db_blkptr)) {
4168 dbuf_release_bp(db);
4170 dbuf_remap(dn, db, tx);
4174 if (parent != dn->dn_dbuf) {
4175 /* Our parent is an indirect block. */
4176 /* We have a dirty parent that has been scheduled for write. */
4177 ASSERT(parent && parent->db_data_pending);
4178 /* Our parent's buffer is one level closer to the dnode. */
4179 ASSERT(db->db_level == parent->db_level-1);
4181 * We're about to modify our parent's db_data by modifying
4182 * our block pointer, so the parent must be released.
4184 ASSERT(arc_released(parent->db_buf));
4185 zio = parent->db_data_pending->dr_zio;
4187 /* Our parent is the dnode itself. */
4188 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
4189 db->db_blkid != DMU_SPILL_BLKID) ||
4190 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
4191 if (db->db_blkid != DMU_SPILL_BLKID)
4192 ASSERT3P(db->db_blkptr, ==,
4193 &dn->dn_phys->dn_blkptr[db->db_blkid]);
4197 ASSERT(db->db_level == 0 || data == db->db_buf);
4198 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
4201 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
4202 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
4203 db->db.db_object, db->db_level, db->db_blkid);
4205 if (db->db_blkid == DMU_SPILL_BLKID)
4207 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
4209 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
4213 * We copy the blkptr now (rather than when we instantiate the dirty
4214 * record), because its value can change between open context and
4215 * syncing context. We do not need to hold dn_struct_rwlock to read
4216 * db_blkptr because we are in syncing context.
4218 dr->dr_bp_copy = *db->db_blkptr;
4220 if (db->db_level == 0 &&
4221 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
4223 * The BP for this block has been provided by open context
4224 * (by dmu_sync() or dmu_buf_write_embedded()).
4226 abd_t *contents = (data != NULL) ?
4227 abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
4229 dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy,
4230 contents, db->db.db_size, db->db.db_size, &zp,
4231 dbuf_write_override_ready, NULL, NULL,
4232 dbuf_write_override_done,
4233 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
4234 mutex_enter(&db->db_mtx);
4235 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
4236 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
4237 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
4238 mutex_exit(&db->db_mtx);
4239 } else if (db->db_state == DB_NOFILL) {
4240 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
4241 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
4242 dr->dr_zio = zio_write(zio, os->os_spa, txg,
4243 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
4244 dbuf_write_nofill_ready, NULL, NULL,
4245 dbuf_write_nofill_done, db,
4246 ZIO_PRIORITY_ASYNC_WRITE,
4247 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
4249 ASSERT(arc_released(data));
4252 * For indirect blocks, we want to setup the children
4253 * ready callback so that we can properly handle an indirect
4254 * block that only contains holes.
4256 arc_write_done_func_t *children_ready_cb = NULL;
4257 if (db->db_level != 0)
4258 children_ready_cb = dbuf_write_children_ready;
4260 dr->dr_zio = arc_write(zio, os->os_spa, txg,
4261 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
4262 &zp, dbuf_write_ready, children_ready_cb,
4263 dbuf_write_physdone, dbuf_write_done, db,
4264 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);